prediction of rainfall time series using modular
TRANSCRIPT
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
1
Prediction of Rainfall Time Series Using Modular Artificial Neural Networks Coupled 1
with Data Preprocessing Techniques 2
C. L. Wu1,2, K. W. Chau1,*, and C. Fan3 3
1 Dept. of Civil and Structural Engineering, Hong Kong Polytechnic University, 4
Hung Hom, Kowloon, Hong Kong, People’s Republic of China 5
6
2 Changjiang Institute of Survey, Planning, Design and Research, 7
Changjiang Water Resources Commission, 8
430010, Wuhan, HuBei, People’s Republic of China 9
10
3 Department of Civil Engineering, Ryerson University, 11
350 Victoria Street, Toronto, Ontario, Canada, M5B 2K3 12
13
*Email: [email protected] 14
15
ABSTRACT 16
This study is an attempt to seek a relatively optimal data-driven model for rainfall forecasting from three aspects: 17
model inputs, modeling methods, and data preprocessing techniques. Four rain data records from different 18
regions, namely two monthly and two daily series, are examined. A comparison of seven input techniques, 19
either linear or nonlinear, indicates that linear correlation analysis (LCA) is capable of identifying model inputs 20
reasonably. A proposed model, modular artificial neural network (MANN), is compared with three benchmark 21
models, viz. artificial neural network (ANN), K-nearest-neighbors (K-NN), and linear regression (LR). 22
Prediction is performed in the context of two modes including normal mode (viz., without data preprocessing) 23
and data preprocessing mode. Results from the normal mode indicate that MANN performs the best among all 24
four models, but the advantage of MANN over ANN is not significant in monthly rainfall series forecasting. 25
Under the data preprocessing mode, each of LR, K-NN and ANN is respectively coupled with three data 26
This is the Pre-Published Version.
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
2
preprocessing techniques including moving average (MA), principal component analysis (PCA), and singular 27
spectrum analysis (SSA). Results indicate that the improvement of model performance generated by SSA is 28
considerable whereas those of MA or PCA are slight. Moreover, when MANN is coupled with SSA, results 29
show that advantages of MANN over other models are quite noticeable, particularly for daily rainfall 30
forecasting. Therefore, the proposed optimal rainfall forecasting model can be derived from MANN coupled 31
with SSA. 32
KEYWORDS 33
Rainfall prediction, Modular artificial neural network, Moving Average, Principal component analysis, Singular 34
spectral analysis, Fuzzy C-Means clustering, K-nearest-neighbors 35
36
1. Introduction 37
Accurate and timely rainfall forecasting is crucial for reservoir operation and flooding 38
prevention because it can provide an extension of lead-time of the flow forecasting, larger 39
than the response time of the watershed, in particular for small and medium-sized 40
mountainous basins. 41
Many studies have been conducted for the quantitative precipitation forecasting using 42
diverse techniques including numerical weather prediction models and remote sensing 43
observations (Yates et al., 2000; Ganguly and Bras, 2003; Sheng, et al., 2006; Doomede et al., 44
2008), statistical models (Chu and He, 1995; Chan and Shi, 1999; DelSole and Shukla, 2002; 45
Munot, 2007; Li and Zeng, 2008; Nayagam et al., 2008), chaos theory-based approach 46
(Jayawardena and Lai, 1994), K-nearest-neighbor (K-NN) method (Toth et al, 2000), and soft 47
computing methods including artificial neural network (ANN), support vectors regression 48
(SVR) and fuzzy inference system (Venkatesan et al., 1997; Silverman and Dracup, 2000; 49
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
3
Toth et al., 2000; Pongracz et al., 2001; Sivapragasam et al., 2001; Brath et al., 2002; Lin and 50
Chen, 2005; Chattopadhyay and Chattopadhyay, 2007; Guhathakurta, 2008; Lin et al., 2009). 51
Venkatesan et al. (1997) employed ANN to predict all India summer monsoon rainfall with 52
different meteorological parameters as model inputs. Toth et al. (2000) applied three data-53
driven models, auto-regressive moving average, ANN and K-NN, to short-term rainfall 54
predictions. Results showed that ANN performed the best in terms of the accuracy of runoff 55
forecasting when the predicted rainfalls by the three models were used as inputs of a rainfall-56
runoff model. Pongracz et al. (2001) applied fuzzy inference to monthly rainfall prediction. 57
Chattopadhyay and Chattopadhyay (2007) constructed an ANN model to predict monsoon 58
rainfall in India depending on the rainfall series alone. 59
Recently, the concept of coupling different models has attracted more attention in 60
hydrologic forecasting. They can be broadly categorized into ensemble models and modular 61
(or hybrid) models. The basic idea behind ensemble models is to build several different or 62
similar models for the same process and to integrate them together (Shamseldin et al., 1997; 63
Shamseldin and O’Connor, 1999; Xiong et al., 2001; Abrahart and See, 2002; Kim et al, 64
2006). For example, Xiong et al. (2001) used a Takagi-Sugeno-Kang fuzzy technique to 65
couple several conceptual rainfall-runoff models. Coulibaly et al. (2005) employed an 66
improved weighted-average method to coalesce forecasted daily reservoir inflows from K-67
NN, conceptual model and ANN. Kim et al. (2006) investigated five ensemble methods for 68
improving stream flow prediction. 69
Physical processes in rainfall and/or runoff are generally composed of a number of 70
sub-processes. Their accurate modeling by building of a single global model is sometimes not 71
possible (Solomatine and Ostfeld, 2008). Modular models were therefore proposed where 72
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
4
sub-processes were first of all identified and then separate models (also called local or expert 73
model) were established for each of them (Solomatine and Ostfeld, 2008). Different modular 74
models were proposed depending on the soft or hard splitting of training data. Soft splitting 75
means the dataset can be overlapped and the overall forecasting output is the weighted-76
average of each local model (Zhang and Govindaraju, 2000; Shrestha and Solomatine, 2006; 77
Wu et al., 2008). Zhang and Govindaraju (2000) examined the performance of modular 78
networks in predicting monthly discharges based on the Bayesian concept. Wu et al. (2008) 79
employed a distributed SVR for daily river stage prediction. On the contrary, there is no 80
overlap of data in the hard splitting and the final forecasting output is explicitly from only 81
one of local models (See and Openshaw, 2000; Hu et al., 2001; Solomatine and Xue, 2004; 82
Sivapragasam and Liong, 2005; Jain and Srinivasulu, 2006; Wang et al., 2006; Corzo and 83
Solomatine, 2007; Lin and Wu, 2009). Hu et al. (2001) developed a range-dependent network 84
which employs a number of multilayer perceptron neural networks to model the river flow in 85
different flow bands of magnitude (e.g. high, medium and low). Their results indicated that 86
the range-dependent network performed better than the conventional global ANN. 87
Solomatine and Xue (2004) used M5 model trees and neural networks in a flood-forecasting 88
problem. Sivapragasam and Liong (2005) divided the flow range into three regions, and 89
employed different SVR models to predict daily flows in high, medium and low regions. 90
Wang et al. (2006) used a crisp modular ANN to make soft or crisp predictions for validation 91
data where each local network was trained using the subsets achieved by either a threshold 92
discharge value or a clustering of input spaces. Lin and Wu (2009) proposed a hybrid ANN 93
model for event-based hourly rainfall prediction where self-organizing map networks are 94
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
5
used for data cluster analysis and multilayer conceptron networks are employed to serve each 95
cluster to construct mapping between input and output. 96
A hydrological time series can be actually regarded as an integration of stochastic (or 97
random) and deterministic components (Salas et al., 1985). Once the stochastic (noise) 98
component is appropriately eliminated, the deterministic component can then be easily 99
modeled. For the purpose of cleaning hydrological series, many data preprocessing 100
techniques, including Principal component analysis (PCA), wavelet analysis (WA), and 101
singular spectrum analysis (SSA), have been employed in hydrology field by researchers 102
(Sivapragasam et al., 2001; Marques et al., 2006; Hu et al., 2007; Partal and Kişi, 2007; 103
Sivapragasam et al., 2007;Wu et al., 2009). Hu et al. (2007) employed PCA as an input data 104
preprocessing tool to improve the prediction accuracy of the rainfall-runoff neural network 105
models. The use of WA to improve rainfall forecasting was conducted by Partal and Kişi 106
(2007). Their results indicated that WA was promising. SSA has also been recognized as an 107
efficient preprocessing algorithm to avoid the effect of discontinuous or intermittent signals, 108
coupled with neural networks (or similar approaches) for time series forecasting (Lisi et al., 109
1995; Sivapragasam et al., 2001; Baratta et al., 2003). For example, Lisi et al. (1995) applied 110
SSA to extract the significant components in their study on southern oscillation index time 111
series and used ANN for prediction. They reconstructed the original series by summing up 112
the first “p” significant components. Sivapragasam et al. (2001) proposed a hybrid model of 113
support vector machine (SVM) and SSA for rainfall and runoff predictions. The hybrid 114
model resulted in a considerable improvement in the model performance in comparison with 115
the original SVM model. A comparison between WA and SSA in Wu et al. (2009) indicated 116
that SSA performed better than WA. In addition, moving average (MA) is used for data 117
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
6
preprocessing to improve the performance of ANN by de Vos and Rientjes (2005). They 118
argued that one of reasons on lagged predictions of ANN was due to the use of previous 119
observed data as ANN inputs and suggested that an effective solution was to obtain new 120
model inputs by MA over the original data series. 121
In this paper, one of the main purposes is to develop a modular ANN (MANN) 122
coupled with appropriate data-preprocessing techniques to improve the accuracy of rainfall 123
forecasting. MANN consists of three local models which are associated with three subsets 124
clustered by the fuzzy C-means (FCM) clustering method. To evaluate MANN, LR, K-NN 125
and ANN are employed for comparison. ANN is first used to choose the best model inputs by 126
seven candidate model inputs techniques. Once all forecasting models are established, three 127
data-preprocessing methods (i.e., MA, PCA, and SSA) can be examined. To ensure wider 128
application of the conclusions, four cases consisting of two monthly rainfall series and two 129
daily rainfall series from India and China, are investigated. The remaining part is structured 130
as follows. Methodology is detailed in Section 2 where case studies are first described, and 131
then data-preprocessing techniques and forecasting models are introduced. Section 3 presents 132
modeling methods and their applications to four rainfall series. The optimal model input 133
method and the best data preprocessing can be identified. In Section 4, principal results are 134
shown along with relevant discussions. The last section presents main conclusions. 135
2. Methodology 136
2.1 Study Area and Data 137
Two daily mean rainfall series from Daning and Zhenshui river basins of China, and 138
two monthly mean rainfall series from India and Zhongxian of China, are analyzed. 139
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
7
The Daning River, a first-order tributary of the Yangtze River, is located at the 140
northeastern side of Chongqing city. The daily rainfall data from Jan. 1, 1988 to Dec. 31, 141
2007 were measured at six raingauges located at the upstream of the study basin (Figure 1). 142
The upstream part is controlled by Wuxi hydrology station, with a drainage area of around 2 143
000 km2. The mean areal rainfall series is calculated by the Thiessen polygon method 144
(hereafter the averaged rainfall series is referred to as Wuxi). 145
The Zhenshui basin is located at the northern side of Guangdong Province and 146
adjoined by Hunan Province and Jianxi Province. The basin belongs to a second-order 147
tributary of the Pearl River and has an area of 7,554 km2. The daily rainfall time series of 148
Zhenwan raingauge was collected between January 1, 1989 and December 31, 1998 149
(hereafter the averaged rainfall series is referred to as Zhenwan). 150
The all Indian average monthly rainfall is estimated from area-weighted observations 151
at 306 land stations uniformly distributed over India. The data, with period spanning from 152
January 1871 to December 2007, are available at the website http://www.tropmet.res.in run 153
by the Indian Institute of Tropical Meteorology. 154
The other monthly rainfall series is from Zhongxian raingauge which is located at 155
Chongqing city, China. The catchment containing this raingauge belongs to a first-order 156
tributary of the Yangtze River. The monthly rainfall data were collected from January 1956 157
to December 2007. 158
Figure 2 shows hyetographs of four rainfall series. A linear fit to each hyetograph is 159
denoted by the dashed line. All series appear stationary at least in a weak sense since these 160
linear fits are close to horizontal. 161
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
8
In this study, each of data series is partitioned into three parts as training set, cross-162
validation set and testing set. The training set serves the model training and the testing set is 163
used to evaluate the performances of models. The cross-validation set has dual functions: one 164
is to implement an early stopping approach in order to avoid overfitting of the training data 165
and another is to select some best predictions from a number of ANN’s runs. In the present 166
study, 10 best predictions are selected from a total of 20 ANN’s runs. The same data partition 167
is adopted for each series: the first half of the entire data as training set and the first half of 168
the remaining data as cross-validation set and the other half as testing set. 169
Table 1 presents pertinent information about watersheds and some descriptive 170
statistics of the original data and three data subsets, including mean (µ), standard deviation 171
(Sx), coefficient of variation (Cv), skewness coefficient (Cs), minimum (Xmin), and maximum 172
(Xmax). As shown in Table 1, the training set cannot fully include the cross-validation or 173
testing data. Owing to the weak extrapolation ability of ANN, all data are scaled to the 174
interval [-0.9, 0.9] instead of [-1, 1] whilst hyperbolic tangent sigmoid functions are 175
employed as transfer functions in hidden and output layers. 176
2.2 Data preprocessing techniques 177
(1) MA 178
MA smoothes data by replacing each data point with the average of the k 179
neighboring data points, where k may be termed the length of memory window. The method 180
is based on the idea that any large irregular component at any point in time will exert a 181
smaller effect if we average the point with its immediate neighbors (Newbold et al., 2003). 182
The equally weighted MA is the most commonly-used, in which each value of the data 183
carries the same weight in the smoothing process. There are three types of moving modes 184
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
9
including centering, backward and forward. In a forecasting scenario, only the backward 185
mode is used since the other two modes may necessitate future observed values. For a time 186
series 1 2, , , Nx x x , when the backward moving mode is adopted (Lee et al., 2000), the 187
-k term unweighted moving average *ty is written as 188
1*
0
k
t t iiy y k
(1) 189
where , ,t k N . The choice of the window length k is by a trial and error procedure with 190
a minimization of the loss of the objective function. 191
(2) PCA 192
PCA was first introduced by Pearson (1901) and developed independently by 193
Hotelling (1933), and has now well entrenched as an important technique in data analysis. 194
The central idea is to reduce the dimensionality of a data set consisting of a large number of 195
interrelated variables, while retaining as much as possible of the variation present in the data 196
set. The PCA approach uses all of the original variables to obtain a smaller set of principal 197
components (PCs) which can be used to approximate the original variables. PCs are 198
uncorrelated and are ordered so that the first few retain most of the variation present in the 199
original set. 200
Consider a data matrix X which has n rows (observations) and p column (variables). 201
Let the covariance matrix of X be , where cov( ) ( )TE X X X . The linear transformed 202
orthogonal matrix Z is presented as 203
Z XA (2) 204
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
10
where Z is the PCs with elements ( ,i j ) of thi observation and thj principal component; A 205
is a ( p p ) matrix with eigenvector elements of the covariance of X , and having 206
T T A A AA I . 207
Because matrix TX X is real and symmetric, it can be expressed as T TX X AΛA 208
where Λ is a diagonal matrix whose nonnegative entries are the eigenvalues (λ , 1, ,i i p ) 209
of TX X . The total variance of the data matrix X is represented as 210
1
trace( ) trace( ) trace( )= λp
ii
TAΛA Λ (3) 211
On the other hand, the covariance matrix of principal components Z is expressed as 212
cov( ) ( ) ( )T T TE E Z Z Z A X XA Λ (4) 213
1
trace( ) trace( )= λp
ii
Z Λ (5) 214
Therefore, the total variance of the data matrix X is identical to the total variance after PCA 215
transformation Z . 216
The solution of PCA, using singular value decomposition (SVD) or determinants of 217
the covariance matrix of X , can provide the eigenvectors A with their 218
eigenvalues, λ , 1, ,i i p , representing the variance of each component after PCA 219
transformation. If the eigenvalues are ordered by 1 2 3λ λ λ λ 0p , the first few PCs 220
can capture most of the variance of the original data while the remaining PCs mainly 221
represent the noise in the data. The percentage of total variance explained by the first thm 222
PCs is 223
1 1
λ λ 100%pm
i ii i
V
(6) 224
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
11
The higher is the selection of the total data variance, V , the better the properties of the data 225
matrix are preserved. For the sake of the reduction of dimensionality, a small number of PCs 226
are selected, though most of the data variance in selected components still remain. If the 227
transformation is to prevent the collinearity of regression variables, the selected component 228
number m in Eq. (6) can be set for a higher total variance, such as 95% 99%V (Hsu et 229
al., 2002). 230
The original data matrix A can be reconstructed by a reverse operation of Eq. (2) as 231
TX ZA (7) 232
By choosing suitable m ( p ) PCs from Z and accompanying m eigenvectors from A , the 233
original data can be filtered. 234
(3) SSA 235
According to Golyandina et al. (2001), the basic SSA consists of two stages: 236
decomposition and reconstruction. The decomposition stage involves two steps: embedding 237
and SVD; the reconstruction stage also comprises two steps: grouping and diagonal 238
averaging. Consider a real-valued time series 1 2, , , NF x x x of length ( 2)N . Assume 239
that the series is a nonzero series, viz. there exists at least one i such that 0ix . Four steps 240
are briefly presented as follows. 241
1st step: embedding 242
The embedding procedure maps the original time series to a sequence of multi-243
dimensional lagged vectors. Let L be an integer (window length), 1 L N , and be the 244
delayed time at a multiple of the sampling period. The embedding procedure forms 245
( 1)n N L lagged vectors 2 ( 1), , , ,T
i i i i i Lx x x x X , where R Li X , and 246
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
12
1,2, ,i n . The ‘trajectory matrix’ of the time series is denoted by [ ]1 i n X X XX 247
having lagged vectors as its columns. In other words, the trajectory matrix is 248
1 2 3
1 2 3
1 2 2 2 3 2 2
1 ( 1) 2 ( 1) 3 ( 1)
n
n
n
L L L N
x x x xx x x xx x x x
x x x x
X
(8) 249
If 1 , the matrix X is called Hankel matrix since it has equal elements on the 250
‘diagonals’ where the sum of subscripts of row and column is equal to constant. If 1 , the 251
equal elements in X are not definitely in the ‘diagonals’. 252
2nd step: SVD 253
Let TS XX , 1 2λ ,λ , ,λL denote the eigenvalues of S taken in the decreasing order 254
of magnitude ( 1 2 3λ λ λ λ 0L ) and 1 2, , , LU U U denote the orthonormal system of 255
the eigenvectors of the matrix S corresponding to these eigenvalues. If we denote 256
λTi i i iV UX ( 1, ,i L ) (equivalent to the thi eigenvector of TX X ), then the SVD of the 257
trajectory matrix X can be written as 258
1 L X X X (9) 259
where λ Ti i i i U VX . The matrices iX have rank 1; therefore they are elementary matrices. 260
The collection ( λ , ,i i iU V ) is called the thi eigentriple of the SVD. Note that iU and iV are 261
also the thi left and right singular vectors of X , respectively. 262
3rd step: grouping 263
The purpose of this step is to identify appropriately the trend component, oscillatory 264
components with different periods, and structureless noises by grouping components. This 265
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
13
step can be skipped if one does not want to precisely extract hidden information by 266
regrouping and filtering of components. 267
The grouping procedure partitions the set of indices {1, , }L into m disjoint subsets 268
1, , mI I , so that the elementary matrix in Eq. (9) is regrouped into m groups. Let 269
1{ , , }pI i i . Then the resultant matrix IX corresponding to the group I is defined as 270
1 pI i i X X X . These matrices are computed for 1, , mI I . By substituting into the 271
expansion (9), one obtains the new expansion 272
1 mII X X X (10) 273
The procedure of choosing the sets 1, , mI I is called the eigentriple grouping. 274
4th step: Diagonal averaging 275
The last step in the basic SSA is the transformation of each resultant matrix of the 276
grouped decomposition (10) into a new series of length N . The diagonal averaging is to find 277
equal elements in the resultant matrix and then to generate a new element by averaging over 278
them. The new element has the same position (or index) as the corresponding elements in the 279
original series. As mentioned in the step 1, the concept of ‘diagonal’ is not true for 1 . 280
Regardless of the value of being larger than or equal to 1, the principle of reconstruction is 281
the same. For 1 , the diagonal averaging can be carried out by the formula recommended 282
by Golyandina et al. (2001). Let Y be a ( L n ) matrix with elements ijy , 1 i L , 1 j n . 283
Let * min( , )L L n , * max( , )n L n and ( 1)N n L . Let *ij ijy y if L n and *
ij jiy y
284
otherwise. Diagonal averaging transfers matrix Y to a series 1 2{ , , , }Ny y y by the following 285
equation 286
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
14
*
*
*
* *, - 1
1
* * *, - 1*
1
- 1* *
, - 1- 1
11
1
1
- 1
k
m k mm
L
k m k mm
N K
m k mm k K
y for k Lk
y y for L k KL
y for L k NN k
(11) 287
Eq. (11) corresponds to the averaging of the matrix elements over the ‘diagonals’ 1i j k . 288
The diagonal averaging, applied to a resultant matrix IkX , produces a N length series kF , 289
and thus the original series F is decomposed into the sum of m series: 290
1 mF F F (12) 291
As mentioned above, these reconstructed components (RCs) can be associated with the trend, 292
oscillations or noise of the original time series with proper choices of L and the sets of 293
1, , mI I . Certainly, if the third step (namely, grouping) is skipped, F can be decomposed 294
into L RCs. 295
2.3 Forecasting models 296
This section describes four candidate forecasting models. They are LR, K-NN, ANN 297
and MANN. They are usually called data-driven models because they capture the mapping 298
between input (e.g. antecedent rainfall) and output variables (forecasted rainfall) without 299
directly considering the physical laws that underlie the mechanism of rainfall (or 300
precipitation). These models are purely based on the information retrieved from the collected 301
rainfall data. 302
(1) Construction of input/output pairs 303
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
15
Let 1 2, , , Nx x x stand for a rainfall time series. It can be reconstructed into a series 304
of delay vectors as t 2 ( 1), , , ,t t t t mx x x x X , where t RmX , is the delay time as a 305
multiple of the sampling period and m is the embedded dimension. Suppose that the rainfall 306
( 1)t T mx at T -step lead is related to the vector tX , the available historical data may be 307
summarized into a set of pairs as t ( 1), : 1, ,t T mx t n X , where n stands for the number of 308
pairs, and ( 1)n N m . 309
The functional relationship between the input vector tX at time t and the predicted 310
output ( 1)Ft T mx
at time t T can be written as follows: 311
( 1) ( )Ft T m t tx f e X (13) 312
where te is a typical noise term, ( 1)Ft T mx
is the prediction of ( 1)t T mx , and ( )f is the 313
mapping function. The difference of various data-driven forecasting models used in the 314
current study relies on the way of approximating ( )f once model inputs are attained with 315
the appropriate selection of ( , m ). 316
(2) LR 317
The linear regression model herein is actually called stepwise linear regression (SLR) 318
model because the forward stepwise regression is used to determine optimal input variables. 319
The basic idea of SLR is to start with a function that contains the single best input variable 320
and to subsequently add potential input variables to the function one at a time in an attempt to 321
improve the model performance. The order of addition is determined by using the partial 322
-F test values to select which variable should enter next. The high partial -F value is 323
compared to a (selected or default) -F to-enter value. After a variable has been added, the 324
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
16
function is examined to see if any variable should be deleted. Interested readers are referred 325
to Draper and Smith (1998) and McCuen (2005) for more details. 326
(3) K-NN 327
The prediction of ( 1)t T mx by the K-NN method is formulated as: 328
( 1) ( 1)( , )
1Ft T m t T m
t S n
x xK
X
(14) 329
where S( , n)X denotes the set of indices t of the K nearest neighbors to the feature vector 330
(n)X . The meaning of “nearest neighbors” is generally interpreted in a Euclidean sense. 331
Therefore, if i belongs to S( , n)X and j is not in S( , n)X , then according to Euclidean 332
distance - -n i n jX X X X . Intuitively speaking, the forecast ( 1)Ft T mx in Eq. (14) is the 333
sample average of output rainfall of the K nearest neighbors to (n)X . Obviously, a key task 334
is to determine the parameter K in the K-NN method. 335
(4) ANN 336
The multilayer perceptron network is by far the most popular ANN paradigm, which 337
usually uses the technique of error back propagation to train the network configuration. The 338
architecture of the ANN consists of a number of hidden layers and a number of neurons in 339
the input layer, hidden layers and output layer. ANNs with one hidden layer are commonly 340
used in hydrologic modeling (Dawson and Wilby, 2001; de Vos and Rientjes, 2005) since 341
these networks are considered to provide enough complexity to accurately simulate the 342
nonlinear-properties of the hydrologic process. Based on Eq. (13), the ANN forecasting 343
model is formulated as 344
( 1) 0 ( 1)1 1
( , , , , ) ( )j
h mF outt T m t ji t i j
j i
x f w m h w w x
X (15) 345
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
17
where denotes transfer functions; jiw are the weights defining the link between the ith 346
node of the input layer and the jth node of the hidden layer; j are biases associated to the 347
jth node of the hidden layer; j
outw are the weights associated to the connection between the 348
jth node of the hidden layer and the node of the output layer; and 0 is the bias at the output 349
node. To apply Eq. (15) to rainfall predictions, appropriate training algorithm is required to 350
optimize w and . 351
(5) MANN 352
To construct MANN, the training data have to be divided into several clusters 353
according to cluster analysis techniques, and then each single model is applied to each cluster. 354
The FCM clustering technique is adopted in the present study (e.g., Bezdek, 1981, Wang et 355
al., 2006). It is able to generate either soft or crisp clusters. ANN (or similar techniques) is 356
unable to extrapolate beyond the range of the data used for training. Otherwise, poor 357
forecasts or predictions can be expected when a new input data is outside the range of those 358
used for training. Hard forecasting is, therefore, taken into consideration in this study. 359
Figure 3 displays the schematic diagram of MANN where the training data is 360
partitioned into three clusters which are based on an assumption that three magnitudes of 361
rainfall (i.e., low, medium, and high) may be derived from different mechanisms. According 362
to this flow chart, once input-output pairs are obtained, they are first split into three subsets 363
by the FCM technique, and then each subset is approximated by a single ANN. The final 364
output of the modular model results directly from the output of one of three local models. 365
2.4 Implementation framework of rainfall forecasting 366
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
18
Figure 4 illustrates the implementation framework of rainfall forecasting where four 367
prediction models can be conducted in two modes: without/with three data preprocessing 368
methods (dashed box). These acronyms in the column of “methods for model inputs” 369
represent seven methods to determine model inputs: LCA (linear correlation analysis, 370
Sudheer et al., 2002), AMI (average mutual information, Fraser and Swinney, 1986), PMI 371
(partial mutual information, May et al., 2008), FNN (false nearest neighbors, Kennel et al., 372
1992), CI (correlation integral, Theiler, 1986), SLR, and MOGA (ANN based on multi-373
objective genetic algorithm, Giustolisi and Simeone, 2006). 374
2.5 Evaluation of model performances 375
The Pearson’s correlation coefficient (r) or the coefficient of determination (R2 = r2), 376
have been identified as inappropriate measures in hydrologic model evaluation by Legates 377
and McCabe (1999). The coefficient of efficiency (CE) (Nash and Sutcliffe, 1970) is a good 378
alternative to r or R2 as a “goodness-of-fit” or relative error measure in that it is sensitive to 379
differences in the observed and forecasted means and variances. Legates and McCabe (1999) 380
also suggested that a complete assessment of model performance should include at least one 381
absolute error measure (e.g., root mean square error (RMSE)) as necessary supplement to a 382
relative error measure. Besides, the Persistence Index (PI) (Kitanidis And Bras, 1980) was 383
adopted here for the purpose of checking the prediction lag effect. Three measures are 384
therefore used in this study. They are listed below. 385
2 2
1 1
ˆCE 1- ( - ) / ( - )n n
i i ii i
y y y y
(16) 386
2
1
1ˆRMSE ( - )
n
i iiy y
n (17) 387
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
19
2 2-
1 1
ˆPI 1- ( - ) / ( - )n n
i i i i li i
y y y y
(18) 388
In these equations, n is the number of observations, ˆiy stands for the forecasted flow, 389
iy represents the observed flow, y denotes the average observed flow, and i ly is the flow 390
estimate from a so-call persistence model (or termed naïve model) that basically takes the last 391
flow observation (at time i minus the lead time l ) as a prediction. CE and PI values of 1 392
stands for perfect fits. A small value of PI may imply occurrence of lagged prediction. 393
3. Applications of Models 394
3.1 Determination of model inputs 395
ANN, equipped with the Levernberg-Marquardt (L-M) training algorithm and 396
hyperbolic tangent sigmoid transfer functions, is used as the benchmark model to examine 397
aforementioned seven model input methods in terms of RMSE. Depending on the simplified 398
algorithm from Yu et al. (2000) (downloaded at http://small.eie.polyu.edu.hk/), the four 399
rainfall series are identified as non-chaotic since the correlation dimension does not display 400
the property of convergence, in particular, for daily rainfall series. Results from remaining six 401
methods are presented in Table 2. These results are based on one step lead prediction and let 402
t+1X be the target value at one-step prediction horizon. It can be seen from RMSE that most 403
of these methods tend to be mutually alternative because their RMSE are close. Owing to the 404
convenience of operation, the LCA method is preferred in this study. Furthermore, Figure 5 405
shows identification of effective inputs in Table 2 for the LCA method. Taking Wuxi and 406
Zhenwan as examples, model inputs should take the previous 5-day and 7-day rainfalls for 407
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
20
them respectively because the partial auto-correlation function (PACF) value decays within 408
the confidence band around these time lags. 409
3.2 Identification of models 410
The model identification is to determine the structure of a candidate model by using 411
training data to optimize relevant parameters of model control once model inputs have been 412
obtained. The LR model is built by the SLR technique. In terms of one step prediction 413
(viz., 1T ), input variables can be found in Table 2. For example, the LR model for Wuxi 414
can be expressed as 415
1 1 2 4 7 110.421 -0.043 0.044 0.025 0.036 0.03Ft t t t t t tx x x x x x x (19) 416
With respect to K-NN, the model identification consists in finding the optimal K if the 417
m dimensional input vector is determined. Sugihara and Mary (1990) suggested that the 418
value of K was taken as 1K m . On the other hand, the choice of K should ensure the 419
reliability of the forecasting (Fraser and Swinney, 1986). The check of robustness of 420
1K m in terms of RMSE is presented in Figure 6, where K is in the interval of [2, 40]. 421
Adopting the value of K as 1m seems reasonable for the current study because the 422
difference between its RMSE and the minimum RMSE is only 2.9% for Wuxi, 2.9% for 423
Zhenwan, 2.6% for India, and 2.0% for Zhongxian, respectively. Consequently, the value of 424
K is 6 for Wuxi ( 5m ), 8 for Zhenwan ( 7m ), 13 for India ( 12m ), and 14 for 425
Zhongxian ( 13m ), respectively. 426
Based on Eq. (14), the formula for one-step lead prediction in the context of K-NN 427
can be defined as 428
1 11
1i
Ft t
K
iKx x
(20) 429
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
21
where 1Xit
stands for an observed value associated with a neighbor of the current state. For a 430
T step lead prediction, Eq. (20) becomes 431
1
1i
Ft T t T
K
i
x xK
(21) 432
The identification of ANN structure is to optimize the number of hidden nodes h in 433
the hidden layer with the known model inputs and output. The optimal size h of the hidden 434
layer is found by systematically increasing the number of hidden neurons from 1 to 10 until 435
the network performance on the cross-validation set no longer improves significantly. Based 436
on the L-M training algorithm and hyperbolic tangent transfer functions, the identified 437
configurations of ANN are 5-5-1 for Wuxi, 7-4-1 for Zhenwan, 12-5-1 for India, and 13-3-1 438
for Zhongxian, respectively. The same method is used to identify the structure of MANN, and 439
the only difference is that the identification is repeated three times, with each time being for a 440
local ANN. Consequently, MANN is obtained as 5-5/7/9-1 for Wuxi, 7-4/8/4-1 for Zhenwan, 441
12-3/2/5-1 for India, and 13-1/1/1-1 for Zhongxian, respectively. 442
It is worthwhile to notice that the standardization/normalization of the training data is 443
very crucial in the improvement of the model performance. Two methods can be found in the 444
literature (Dawson and Wilby, 2001; Cannas et al., 2002; Rajurkar et al, 2002; Campolo et al., 445
2003; Wang et al., 2006). The standardization (also termed rescaling in some papers) method, 446
as adopted above for model input determination, is to rescale the training data to [-1, 1], [0, 1] 447
or even more narrow interval depending on what kinds of transfer functions are employed in 448
ANN. The normalization method is to rescale the training data to a Gaussian function with a 449
mean of 0 and unit standard deviation, which is by subtracting the mean and dividing by the 450
standard deviation. When the normalization approach is adopted, ANN uses the linear 451
function (e.g. purelin) instead of the hyperbolic tangent sigmoid transfer function in the 452
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
22
output layer. In addition, some studies have indicated that considerations of statistical 453
principles may improve ANN model performance (e.g. Cheng and Titterington, 1994). For 454
example, the training data was recommended to be normally distributed (Fortin et al., 1997). 455
Sudheer et al. (2002) suggested that the issue of stationarity should be considered in the ANN 456
development because the ANN cannot account for trends and heteroscedasticity in the data. 457
Their results showed that data transformation to reduce the skewness of data was capable of 458
significantly improving the model performance. For the purpose of obtaining better model 459
performance, four data-transformed schemes are examined: 460
Standardizing the raw data (referred to as Std_raw); 461
Normalizing the raw data (referred to as Norm_raw); 462
Standardizing the n-th root transformed data (referred to as Std_nth_root); 463
Normalizing the n-th root transformed data (referred to as Norm_nth_root). 464
Table 3 compares the ANN model performance of the four schemes in terms of 465
RMSE and CE. The Norm_raw scheme is, on the whole, slightly more effective than the 466
Std_raw method. It can also be seen that the effect of the n-th root scheme (3 is taken after 467
trial and error) on the improvement of the performance is basically negligible. Therefore, the 468
Norm_raw scheme is adopted for the later rainfall prediction in the present study. 469
3.3 Rainfall data preprocessing 470
(1) MA 471
The MA operation entails the window length k in Eq. (1) to smooth the raw rainfall 472
data. An appropriate k can be found by systematically increasing k from 1 to 10. The 473
smoothed data is then used to feed into each forecasting model. The targeted value of k 474
corresponds to the optimal model performance in terms of RMSE. 475
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
23
(2) PCA 476
PCA is employed in two ways: one for reduction of the dimensionality or preventing 477
collinearity (depending on Eq. (2)); second for noise reduction by choosing leading 478
components (contributing most of the variance of the original rainfall data) to reconstruct 479
rainfall series (depending on Eq. (7)). The percentage V of total variance (see Eq. (6)) is set 480
at three horizons, 85%, 90%, and 95% for principal component selection. 481
(3) SSA 482
This approach of filtering a time series to retain desired modes of variability is based 483
on the idea that the predictability of a system can be improved by forecasting the important 484
oscillations in time series taken from the system. The general procedure is to filter the 485
original record first and then to build the forecasting model based on the filtered series. To 486
filter the raw rainfall series, the series needs to be decomposed into components with the aid 487
of SSA. The decomposition by SSA requires identifying the parameter pair ( , L ). The value 488
of an appropriate L should be able to clearly resolve different oscillations hidden in the 489
original signal. However, the present study does not require accurately resolving the raw 490
rainfall signal into trends, oscillations, and noises. A rough resolution can be adequate for the 491
separation of signals and noises where some leading eigenvalues should be identified. 492
To select L , a small interval of [3, 10] is examined in the present study. Figure 7 493
shows the relation between singular spectrum (namely, a set of singular values) and singular 494
number L for Wuxi, Zhenwan, India, and Zhongxian. It can be observed that the curve of 495
singular values in each case except for Wuxi tends to level off with the increase of L . 496
Generally, extraction of high-frequency oscillations becomes more difficult with the increase 497
of singular number L (or mode). L is selected empirically by following the criterion that the 498
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
24
singular spectrum can be distinguished markedly under that L . According to this criterion, 499
L is set the value of 7 for India and Zhenwan, 6 for Zhongxian. For Wuxi, since all values in 500
the interval satisfy the criterion, in order to reduce computational load in later filtering 501
operation, L is set a small value of 5. The singular spectrum associated with the selected L is 502
highlighted by the dotted solid line in Figure 7. 503
As regards , Figure 8 presents the results of sensitivity analysis of singular spectrum 504
on the lag time using SSA with the determined L . For daily rainfall series, the singular 505
spectrum can be distinguished only when 1 . In contrast, the singular spectrum is 506
insensitive to in the case of monthly rainfall series. The final parameter pair ( , L ) in SSA 507
are set as (1, 5) for Wuxi, (1, 7) for Zhenwan, (1, 7) for India, (1, 6) for Zhongxian, 508
respectively. 509
3.4 Filtering of RCs 510
The subsequent task is to reconstruct a new rainfall series as model inputs by finding 511
contributing RCs so as to improve the predictability of the rainfall series. There is no 512
practical guide on how to identify a contributing or noncontributing component to the 513
improvement of accuracy of prediction. Two proposed filtering methods, supervised and 514
unsupervised, are herein examined. 515
(1) Supervised filtering (denoted by SSA1) 516
Figure 9 depicts cross-correlation function (CCF) between RCs and the original 517
Zhenwan rainfall series. The last plot in this figure presents the average of CCFs from all 7 518
RCs. The average indicates an overall correlation between input and output at various lags 519
(also termed prediction horizons). The plot of average CCF shows that the best correlation is 520
positive and occurs at lag 1. Among all 7 RCs, RC1 exhibits the best positive correlation with 521
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
25
the original rainfall series. The CCF values for other RCs change alternatively between 522
positive and negative with the increase of the lag. From the perspective of linear correlation, 523
the positive or negative CCF value may indicate that the RC makes a positive or negative 524
contribution to the output of model when the RC is used as the input of model. With the 525
assumption, deleting RCs, which have negative correlations with the model output if the 526
average CCF is positive, may improve the performance of the forecast model. This is the 527
basic idea behind the supervised method. 528
The procedure of the supervised method coupled with ANN is depicted in Figure 10. 529
The aim is to find the optimal p ( L ) RCs from all L RCs for each prediction horizon. The 530
procedure can be summarized into three steps: SSA decomposition, correlation coefficients 531
sorting, and reconstructed components filtering. Operation in each step is bounded by the 532
dashed box. It is worth noting that the filtering method is based on assumption that 533
combination of components with the same sign in CCF ( or -) can strengthen the correlation 534
with the model output. 535
(2) Unsupervised filtering (denoted by SSA2) 536
There are some drawbacks on the supervised method. The salient one is that this 537
method relies on linear correlation analysis, which disregards the existence of nonlinearity in 538
meteorological processes. Also, random combinations among all RCs are not taken into 539
account. To overcome these drawbacks, an unsupervised filtering method (also termed 540
enumeration) is recommended where all input combinations are examined. There are 541
2L combinations for L RCs. The unsupervised method may be computationally intensive if 542
L is large. 543
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
26
4. Results and Discussions 544
This section presents predictions using various models under two types of modes, 545
namely “normal” and “data preprocessing”. The “data preprocessing” mode is separately 546
described by MA, PCA, and SSA. To extend one-step-ahead prediction to multi-step-ahead 547
prediction, a direct multi-step prediction method (by directly having the multi-step-ahead 548
prediction as output, also termed static prediction method) is adopted in this study to perform 549
two- and three-step-ahead predictions. 550
4.1 Forecasting with normal mode 551
Table 4 shows results of three prediction horizons by applying five models including 552
naïve model to each case study. The naïve model is used as the benchmark in which the 553
forecasted value is directly equal to the last observed value (namely, no change). The naïve 554
model presents the poorest forecasting which can be explained by the fact that it is unlikely to 555
capture any dependence relation. From the perspective of rainfall series, the monthly rainfall 556
can be better predicted than the daily rainfall. Generally, a daily rainfall series, in particular 557
in a semi-humid and semi-dry or dry region, tends to be intermittent and discontinuous due to 558
a large number of no rain periods (dry periods). Two global modeling methods, LR and ANN, 559
mainly capture the zero-zero (or similar extreme low-intensity) rainfall patterns in daily 560
rainfall series because the type of pattern is overwhelmingly dominant in the daily rainfall 561
series. As a consequence, poor performance indices in terms of RMSE, CE, and PI can be 562
observed (depicted in Table 4 for Wuxi and Zhenwan). Nevertheless, Table 4 also shows that 563
MANN performs the best in each case study. MANN adopts three local ANN models, one for 564
each cluster generated by FCM, which can better capture the mapping relation than using a 565
single global ANN. It can be noticed that MANN is more effective for daily rainfall series 566
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
27
than monthly rainfall data, which can be because daily rainfall data is more irregular (or non-567
periodic) than monthly rainfall series. The use of K-NN for daily rainfall forecasting is even 568
worse than LR although it employs a local prediction approach. Apart from the issue of the 569
selection of K , the performance of K-NN is also influenced by the similarity of input-output 570
patterns. The smooth monthly rainfall series easily construct similar patterns so that they are 571
well predicted by K-NN. It is worth noting that negative values occasionally appear in the 572
forecasts of ANN or MANN whereas this situation does not happen in the K-NN method. 573
Take Wuxi and India data as representative examples, Figure 11 shows the scatter 574
plots and hyetographs of the results at one-day-ahead prediction of ANN and MANN using 575
the rainfall data of Wuxi, where the hyetograph is plotted in a selected range for better visual 576
inspection. ANN seriously underestimates a number of moderate- and high-intensity rainfalls. 577
The low values of CE and PI demonstrate that time shift between the forecasted and observed 578
rainfall may occur, which is further verified by the hyetograph. MANN improves noticeably 579
the accuracy of forecasting in terms of CE and PI. As shown by the scatter plots, the 580
medium-intensity rainfall can be simulated better by MANN although high-intensity rainfalls 581
(or peak values) are still underestimated. Figure 12 shows scatter plots and hyetographs of 582
results at one-day-ahead prediction of ANN and MANN using the rainfall data of India. It 583
can be seen from hyetograph graphs that both ANN and MANN reproduce well the 584
corresponding observed rainfall data, which is further revealed by the scatter plots with a low 585
dispersion around the exact fit line. 586
Figure 13 shows the analysis of the lag effect between forecasted and observed 587
rainfall series. The value of CCF at zero lag corresponds to the actual performance (i.e. 588
correlation coefficient) of the model. A target lag is associated with the maximum value of 589
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
28
CCF, and is an expression for the mean lag for the forecast. It can be seen from Figure 16 590
that ANN makes fairly obvious lagged predictions for daily rainfall series, and the lag effect 591
can be overcome by MANN. There are 1, 2, and 3 days lag for Wuxi, which are respectively 592
associated with one-, two-, and three-day-ahead forecasting. In contrast, there is no lag effect 593
in monthly rainfall predictions of ANN or MANN. 594
4.2 Forecasting with MA 595
Table 5 presents forecasted results of ANN with the “backward” MA (hereafter 596
referred to as ANN-MA) using the Wuxi rainfall data. The performance indices 597
corresponding to 1k are associated with the normal ANN. Results at each prediction 598
horizon seem to be insensitive to the window length k in view of slight differences among 599
each performance index for k from 1 to 10. Considering the fact that ANNs tend to generate 600
unstable outputs, the influence of MA on the performance of ANN is negligible. Small values 601
of PI also imply that MA cannot eliminate the lagged forecast from ANN. 602
4.3 Forecasting with PCA 603
As mentioned previously, PCA is used in two ways: one (denoted by PCA1) for 604
reduction of dimensionality (also termed principal component regression) and the other one 605
(denoted by PCA2) for noise reduction. Results from PCA1 are presented in Table 6. The 606
scenario of V 100% stands for forecasting using models with the normal mode. Results 607
show that PCA1 cannot improve the model performances in terms of RMSE, CE, and PI, 608
which means that the reduction of dimensionality is unnecessary for the present case studies. 609
Actually, the original inputs are characterized by a low dimension. Table 7 describes the 610
results from PCA2. According to results from LR and K-NN (because results from ANN tend 611
to be unstable), a marginal improvement in the model performances can be observed for the 612
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
29
Wuxi watershed whereas the model performances deteriorate for the India watershed with the 613
decrease of the value of V . 614
4.4 Forecasting with SSA 615
Following the procedure in Figure 10, the supervised filtering (SSA1) using ANN for 616
RCs of Wuxi and India is illustrated in Figure 14. The RMSE associated with the maximum 617
number of p (for instance 5p for Wuxi) represents the performance of ANN with the 618
normal mode. The optimal p corresponds to the minimum RMSE, which can be found by 619
systematically deleting RCs one at a time. Consequently, numbers of chosen optimal p RCs 620
in three forecasting horizons are 3, 2, and 1 for Wuxi, and 1, 3, and 5 for India, respectively. 621
However, the unsupervised filtering method (SSA2) is based on enumeration of combinations 622
of all RCs. Selection of the optimal p RCs cannot be presented in a graphical form. 623
Table 8 shows selected p at various prediction horizons using LR, K-NN, and ANN 624
in conjunction with SSA1 and SSA2. A large amount of information can be extracted from 625
this Table. First of all, a considerable improvement in the model performance is achieved by 626
each forecasting model in conjunction with SSA1 or SSA2, compared with results in Table 4. 627
From the perspective of rainfall series, the accuracy of daily rainfall prediction is improved 628
significantly in comparison to that in the normal mode. Secondly, as expected, results from 629
SSA2 are superior to or at least equivalent to those from SSA1 since the former examines 630
each combination of RCs in search of the optimal p . SSA2 is therefore considered as an 631
efficient and effective method if the number L of RCs is small. For the present four cases, 632
SSA2 method is appropriate due to the small number of RCs. Once L is large, say 40 or 50, 633
SSA1 may be a good alternative where a relative optimal forecasting can be guaranteed. 634
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
30
Additionally, it should be noted that the optimal p are different at three forecast horizons. 635
Finally, Table 8 also shows that, among the three models, ANN performs the best with SSA1 636
or SSA2, which is consistent with results in the normal mode. 637
In the normal mode, MANN has been proved to be superior to ANN, in particular, for 638
daily rainfall forecasting. As an attempt to improve the accuracy of rainfall forecasting, 639
MANN is also coupled with SSA2. Table 9 demonstrates results in terms of RMSE, CE, and 640
PI using MANN compared with those of ANN. Good accuracies of forecasting are made by 641
both MANN and ANN. It can be seen from values of PI that the prediction lag effect is 642
completely eliminated. The model performance does not deteriorate markedly with the 643
increase of the forecasting lead. Results also show that MANN still maintains a salient 644
superiority over ANN in the SSA2 mode for both daily and month rainfall series. 645
One-step lead estimates of MANN and ANN with the help of SSA2 are shown in 646
Figure 15 (Wuxi) and Figure 16 (India) in the form of hyetographs and scatter plots (the 647
former is plotted in a selected range for better visual inspection). Compared with Figure 11, 648
each scatter plot in Figure 15 is closer to the exact line, which means that the daily rainfall 649
process is fitted appropriately. Nevertheless, some peak values still remain mismatched 650
although MANN shows a better ability to capture the peak value than ANN. Regarding the 651
monthly rainfall series, the scatter plots with perfect match of the diagonal indicates that the 652
rainfall process is perfectly reproduced. The representative hyetograph shows that the peak 653
times and peak values are also accurately predicted. 654
Figure 17 presents the correlation analysis between observed and forecasted rainfall 655
from ANN and MANN using the Wuxi and India series, respectively. Compared with Figure 656
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
31
13, the lagged prediction of ANN is completely eliminated by SSA2 since the maximum 657
CCF occurs at zero lag. The larger the CCF at zero lag is, the better the model performance is. 658
4.5 Discussions 659
Some discussions regarding forecasting models and the effects of the SSA technique 660
are made in the following. 661
(1) About the investigation of effects of SSA 662
Figure 18 shows that a large number of zeros and near zeros occur in the original 663
Wuxi rainfall which makes the series discontinuous. Using the intermittent series is difficult 664
to reconstruct similar input patterns for a forecasting model. Thus, depending on those 665
reconstructed input patterns, data-driven models based on pattern training, for example, 666
ANNs, tend to be unfeasible. In contrast, rainfall series preprocessed by SSA becomes 667
smoother where most of zeros are replaced by nonzero values. New input vectors from the 668
reconstructed rainfall series are characterized by better repetition of patterns so that they are 669
easier reproduced. 670
To investigate the influence of SSA on the ANN’s performance, correlation analyses 671
between inputs and output of ANN and ANN-SSA2 are compared using the Wuxi data and 672
are depicted in Figure 19. As the input and output series in ANN are both the raw rainfall 673
series, the cross-correlation analysis is equivalent to the autocorrelation analysis of the raw 674
rainfall series. At all three prediction horizons, cross-correlation coefficients (CCs) between 675
reconstructed inputs by SSA2 and the raw rainfall data are improved significantly at most 676
lags except for the lag of 3 at one-step lead. It should be recalled that model inputs are the 677
previous five rainfall data for Wuxi. The “starting point” in Figure 19 represents the first 678
previous rainfall of the five inputs, and the remaining inputs consist of four points after the 679
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
32
starting point. It can be observed that the CCF value between each new model input and 680
output are far larger than that between the raw model input and output (seen at the same lag). 681
Therefore, the improvement of a model’s performance by the SSA technique may be owing 682
to the enhancement of the mapping relation of model input and output by deleting noises 683
hidden in the raw signals. 684
(2) About parameter L in SSA 685
The parameter L in SSA has a significant impact on the performance of a forecasting 686
model since the optimal p RCs may be different with the change of L when using the same 687
forecasting model at the same prediction horizon. The selection of L in this study is based on 688
the interval of [3, 10] in conjunction with an empirical criterion (namely, a particular L is 689
selected only if the singular spectrum can be distinguished markedly under that L ). To check 690
the robustness of the empirical method, each L in [3, 10] is examined by the LR model with 691
SSA2 using the Wuxi and Zhenwan data and presented in Table 10. As mentioned previously, 692
the target L for Wuxi and Zhenwan are 5 and 7, respectively. The RMSE associated with 693
them at each prediction horizon is highlighted in bold (shown in Table 10). In terms of Wuxi, 694
the difference between the target RMSE and the minimum RMSE at the same prediction 695
horizon is only 9.2% for one-step prediction, 1.4% for two-step prediction, 0.0% for three-696
step prediction, respectively. Regarding Zhenwan, the three values are respectively 5.2%, 697
7.5%, and 0.0%. These changes are slight and cannot influence the conclusions drawn 698
previously. Therefore, the empirical method for the present rainfall data should be 699
appropriate. 700
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
33
5. Conclusion 701
This study suggests the use of modular artificial neural network (MANN) coupled 702
with data preprocessing techniques for improving four rainfall predictions from India and 703
China consisting of two monthly and two daily series. To reasonably evaluate MANN’s 704
performance, three models, LR, K-NN and ANN, are used for the purpose of comparison. In 705
the process of model development, model inputs and data preprocessing techniques are 706
carefully analyzed and discussed. The following conclusions are reached based on this study: 707
1. LCA is regarded as an effective and efficient method among all seven input 708
techniques due to its simplicity of computation and comparable capability of forecasting. 709
2. In the normal mode (without data preprocessing), MANN distinguishes from the 710
other three models for both monthly and daily rainfall series forecasting. Whilst all four 711
models reasonably forecast two monthly rainfall series, only MANN is able to simulate each 712
daily rainfall series without obvious lag effect. 713
3. In the data preprocessing mode, the effect of MA is negligible for the improvement 714
of each forecasting model. 715
4. PCA as a data preprocessing technique is discussed in two forms, i.e. PCA1 for the 716
purpose of dimension reduction, and PCA2 for the purpose of noise reduction. Results show 717
that PCA1 cannot improve model’s performance and PCA2 marginally improve model’s 718
performance. 719
5. Two filtering method, i.e. supervised (SSA1) and unsupervised (SSA2), are 720
examined for SSA when coupled with forecasting models. It can be found that each model 721
achieves considerable improvement in performance with the aid of SSA1 or SSA2. In terms 722
of forecasting models, MANN still outperforms all other models. 723
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
34
6. As far as two filtering methods are concerned, SSA2 tends to be better if the 724
number of raw RCs is small. Otherwise, SSA1 is a good alternative. 725
7. A further discussion reveals that the essence of SSA in improving model 726
performance is to strengthen the mapping relation of model input and output by deleting 727
noises in the raw signal. 728
8. There is still considerable room for improving forecasting of peak values although 729
MANN coupled with SSA has made perfect overall predictions for daily rainfall series. 730
731
Nomenclature 732
ACF Auto Correlation Function 733
AMI Average Mutual Information 734
ANN Artificial Neural Networks 735
CC Cross-correlation Coefficient 736
CCF Cross Correlation Function 737
CE Coefficient of Efficiency 738
CI Correlation Integral 739
FCM Fuzzy C-means 740
FNN False Nearest Neighbor 741
K-NN K-Nearest-Neighbors 742
LCA Linear Correlation Analysis 743
L-M Levenberg-Marquart 744
LR Linear Regression 745
MA Moving Average 746
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
35
MANN Modular Artificial Neural Networks 747
MOGA ANN based on Multi-objective Genetic Algorithm 748
PACF Partial Auto Correlation Function 749
PC Principal Component 750
PCA Principal Component Analysis 751
PMI Partial Mutual Information 752
PI Persistence Index 753
RC Reconstructed Component 754
RMSE Root Mean Square Error 755
SLR Stepwise Linear Regression 756
SSA Singular Spectrum Analysis 757
SVD Singular Value Decomposition 758
SVM Support Vector Machine 759
SVR Support Vectors Regression 760
WA Wavelet Analysis 761
762
References: 763
Abrahart, R.J. and See, L.M. (2002), Multi-model data fusion for river flow forecasting: an 764
evaluation of six alternative methods based on two contrasting catchments. Hydrology and 765
Earth System Sciences, 6(4), 655-670. 766
Baratta et al., Baratta, D., Cicioni, G., Masulli, F. and Studer, L. (2003), Application of an 767
ensemble technique based on singular spectrum analysis to daily rainfall forecasting. Neural 768
Networks, 16, 375-387. 769
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
36
Bezdek, J.C. (1981), Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum 770
Press, New York. 771
Brath, A., Montanari, A., and Toth, E. (2002), Neural networks and non-parametric methods 772
for improving realtime flood forecasting through conceptual hydrological models. Hydrology 773
and Earth System Sciences, 6(4), 627-640. 774
Campolo, M., Andreussi, P. and Soldati, A. (2003), Artificial neural network approach to 775
flood forecasting in the river Arno. Hydrological Sciences Journal, 48(3), 381-398. 776
Cannas, B., Fanni, A., Pintus, M., and Sechi, G. M. (2002), Neural Network Models to 777
Forecast Hydrological Risk. IEEE IJCNN, 2002. 778
Chan, J.C.L., and Shi, J.E. (1999), Prediction of the summer monsoon rainfall over South 779
China. International Journal of Climatology, 19 (11), 1255-1265. 780
Chattopadhyay,S., and Chattopadhyay, G. (2007), Identification of the best hidden layer size 781
for three-layered neural net in predicting monsoon rainfall in India. Journal of 782
Hydroinformatics, 10(2), 181-188. 783
Cheng, B., and Titterington, D. M. (1994), Titterington Neural networks: A review from a 784
statistical perspective. Statistical Science, 9(1), 2-54. 785
Chu, P.S., and He, Y.X. (1994), Long-Range Prediction of Hawaiian Winter Rainfall Using 786
Canonical Correlation-Analysis). International Journal of Climatology, 14(6), 659-669. 787
Corzo, G., and Solomatine, D. (2007), Baseflow separation techniques for modular artificial 788
neural network modelling in flow forecasting. Hydrological Sciences–Journal–des Sciences 789
Hydrologiques, 52(3), 491-507. 790
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
37
Coulibaly, P., Haché, M., Fortin, V., and Bobée, B. (2005), Improving daily reservoir inflow 791
forecasts with model combination. Journal of Hydrologic Engineering, 10(2), 91-99. 792
Dawson, C.W., and Wilby, R.L. (2001), Hydrological Modeling Using Artificial Neural 793
Networks. Progress in Physical Geography, 25(1), 80-108. 794
DelSole, T., and Shukla, J. (2002), Linear prediction of Indian monsoon rainfall. Journal of 795
Climate, 15 (24), 3645-3658. 796
de Vos, N.J. and Rientjes, T.H.M. (2005), Constraints of artificial neural networks for rainfall 797
-runoff modeling: trade-offs in hydrological state representation and model evaluation. 798
Hydrology and Earth System Sciences, 9, 111-126. 799
Diomede, T., Davolio, S., Marsigli, C., Miglietta, M. M., Moscatello, A., Papetti, P., 800
Paccagnella, T., Buzzi, A., and Malguzzi, P. (2008), Discharge prediction based on multi-801
model precipitation forecasts. Meteorology and Atmospheric Physics, 101 (3-4), 245-265. 802
Draper, N. R. and Smith, H. (1998), Applied regression analysis, 3rd ed. New York: Wiley. 803
Fortin, V., Ouarda, T.B.M.J., and Bobe´e, B., (1997), Comment on ‘The use of artificial 804
neural networks for the prediction of water quality parameters’ by H.R. Maier and G.C. 805
Dandy. Water Resources Research, 33 (10), 2423-2242. 806
Fraser, A.M. and Swinney, H.L. (1986), Independent coordinates for strange attractors from 807
mutual information, Physical Review A, 33(2), 1134-1140. 808
Ganguly, A.R., and Bras, R.L. (2003), Distributed quantitative precipitation forecasting 809
(DQPF) using information from radar and numerical weather prediction models. Journal of 810
Hydrometeorology, 4 (6), 1168-1180. 811
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
38
Giustolisi, O., and Savic, D. A. (2006), A symbolic data-driven technique based on 812
evolutionary polynomial regression, Journal of Hydroinformatics, 8(3), 207-222. 813
Golyandina, N. Nekrutkin,V., and Zhigljavsky, A. (2001), Analysis of Time Series Structure: 814
SSA and Related Techniques, Chapman & Hall/CRC. 815
Guhathakurta, P. (2008), Long lead monsoon rainfall prediction for meteorological sub-816
divisions of India using deterministic artificial neural network model. Meteorology and 817
Atmospheric Physics, 101 (1-2), 93-108. 818
Guhathakurta, P., Rajeevan, M., and Thapliyal, V. (1999), Long range forecasting Indian 819
summer monsoon rainfall by a hybrid Principal Component neural network model. 820
Meteorology and atmospheric physics, 71, (3-4), 255-266. 821
Hsu, K.L., Gupta, H.V., Gao, X.G., Sorooshian, S., and Imam, B. (2002), Self-organizing 822
linear output map (SOLO): An artificial neural network suitable for hydrologic modeling and 823
analysis. Water Resources Research, 38(12), 1302,doi:10.1029/2001WR000795. 824
Hu, T.S., Lam, K.C., Ng, S.T., (2001), River flow time series prediction with a range-825
dependent neural network. Hydrological Science Journal, 46 (5), 729-745. 826
Hotelling, H., (1933), Analysis of a complex of statistical variables into principal components. 827
Journal of Educational Psychology, 24, 417-441. 828
Hu, T.S.,Wu, F.Y., and Zhang, X. (2007), Rainfall-runoff modeling using principal 829
component analysis and neural network. Nordic Hydrology, 38(3), 235-248. 830
Jain, A., and Srinivasulu, S. (2006), Integrated approach to model decomposed flow 831
hydrograph using artificial neural network and conceptual techniques. Journal of Hydrology, 832
317, 291-306. 833
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
39
Jayawardena, A.W., and Lai, F. (1994), Analysis and prediction of chaos in rainfall and 834
stream flow time series. Journal of hydrology, 153, 23-52. 835
Kennel, M. B., Brown, R., and Abarbanel, H. D. I. (1992), Determining embedding 836
dimension for phase space reconstruction using geometrical construction. Physical Review A., 837
45(6), 3403-3411. 838
Kim, T., Heo, J. H. and Jeong, C. S. (2006), Multireservoir system optimization in the Han 839
River basin using multi-objective genetic algorithms. Hydrological Processes, 20 (9), 2057-840
2075. 841
Kitanidis, P. K. and Bras, R. L. (1980), Real-time forecasting with a conceptual hydrologic 842
model, 2, applications and results, Water Resources Research, 16 (6), 1034–1044. 843
Lee, C.F., Lee, J.C. and Lee, A.C., (2000), Statistics for business and financial economics 844
(2nd version), World Scientific, Singapore. 845
Legates, D. R., and McCabe, Jr, G. J. (1999), Evaluating the use of goodness-of-fit measures 846
in hydrologic and hydroclimatic model validation, Water Resources. Research, 35(1), 233- 847
241. 848
Li, F., and Zeng, Q.C. (2008), Statistical prediction of East Asian summer monsoon rainfall 849
based on SST and sea ice concentration. Journal of the meteorological society of Japan, 86 850
(1), 237-243. 851
Lin, G.F., and Chen, L.H. (2005), Application of artificial neural network to typhoon rainfall 852
forecasting. Hydrological Processes, 19 (9), 1825-1837. 853
Lin, G.F., Chen, G.R., Wu, M.C., and Chou, Y.C. (2009), Effective forecasting of hourly 854
typhoon rainfall using support vector machines. Water Resources Research, 45, W08440. 855
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
40
doi:10.1029/2009WR007911. 856
Lin, G.F., and Wu, M.C. (2009), A hybrid neural network model for typhoon-rainfall 857
forecasting. Journal of Hydrology, 375 (3-4), 450-458. 858
Lisi, F., Nicolis, and Sandri, M. (1995), Combining singular-spectrum analysis and neural 859
networks for time series forecasting. Neural Processing Letters, 2(4), 6-10. 860
Marques, C.A.F., Ferreira, J., Rocha, A., Castanheira, J., Gonçalves, P., Vaz. N., and Dias, 861
J.M. (2006), Singular spectral analysis and forecasting of hydrological time series. Physics 862
and Chemistry of the Earth, 31, 1172-1179. 863
May, R.J., Maier, H.R., Dandy, G.C., and Fernando, T.M.K. (2008), Non-linear variable 864
selection for artificial neural networks using partial mutual information. Environmental 865
Modeling & Software, 23, 1312-1328. 866
McCuen, R. H. (2005), Hydrologic analysis and design (3rd ed.), Upper Saddle River, NJ: 867
Pearson/Prentice Hall. 868
Munot, A. A., and Kumar, K.K. (2007), Long range prediction of Indian summer monsoon 869
rainfall. Journal of Earth System Science, 116 (1), 73-79. 870
Nash, J. E. and Sutcliffe, J. V. (1970), River flow forecasting through conceptual models part 871
I — A discussion of principles. Journal of Hydrology, 10 (3), 282-290. 872
Nayagam, L.R., Janardanan, R., and Mohan, H.S.R. (2008), An empirical model for the 873
seasonal prediction of southwest monsoon rainfall over Kerala, a meteorological subdivision 874
of India. International Journal of Climatology, 28 (6), 823-831. 875
Newbold, P., Carlson, W. L., and Thorne, B.M. (2003), Statistics for business and economics 876
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
41
(fifth version), Prentice Hall, Upper Saddle River, N.J. 877
Partal, T. and Kişi, Ӧ. (2007), Wavelet and Neuro-fuzzy conjunction model for precipitation 878
forecasting. Journal of Hydrology, 342 (2), 199-212. 879
Pearson, K., (1901), On lines and planes of closest fit to systems of points in space. 880
Philosophical Magazine, 2, 559-572. 881
Pongracz, R., Bartholy, J., and Bogardi, I. (2001), Fuzzy rule-based prediction of monthly 882
precipitation. Physics and Chemistry of the Earth Part B-Hydrology Oceans and Atmosphere, 883
26 (9), 663-667. 884
Rajurkar, M.P., Kothyari, U.C., and Chaube, U.C. (2002), Artificial neural networks for daily 885
rainfall-runoff modeling. Hydrological Sciences-Journal, 47(6), 865-876. 886
Salas, J.D., Delleur, J.W., Yevjevich, V., and Lane, W.L. (eds) (1985), Applied Modeling of 887
Hydrologic Time Series. Water Resources Publications: Littleton, Colorado. 888
See, L., and Openshaw, S. (2000), A hybrid multi-model approach to river level forecasting. 889
Hydrology Science Journal, 45(4), 523-536. 890
Shamseldin, A.Y. and O’Connor, K.M. (1999), A real-time combination method for the 891
outputs of different rainfall-runoff models. Hydrological Sciences Journal, 44(6), 895-912. 892
Shamseldin, A.Y., O'Connor, K.M. and Liang, G.C., (1997), Methods for combining the 893
outputs of different rainfall runoff models, Journal of Hydrology, 197, 203–229. 894
Sheng, C., Gao, S., and Xue, M. (2006), Short-range prediction of a heavy precipitation event 895
by assimilating Chinese CINRAD-SA radar reflectivity data using complex cloud analysis. 896
Meteorology and Atmospheric Physics, 94 (1-4), 167-183. 897
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
42
Shrestha, D.L., and Solomatine, D.P. (2006), Machine learning approaches for estimation of 898
prediction interval for the model output. Neural Networks, 19(2), 225-235. 899
Silverman, D., and Dracup, J.A. (2000), Artificial neural networks and long-range 900
precipitation in California. Journal of Applied Meteorology 31 (1), 57-66. 901
Sivapragasam, C., and Liong, S. Y. (2005), Flow categorization model for improving 902
forecasting. Nordic Hydrology 36 (1), 37-48. 903
Sivapragasam, C., Liong, S.Y. and Pasha, M.F.K. (2001), Rainfall and runoff forecasting with 904
SSA-SVM approach. Journal of Hydroinformatics, 3(7), 141-152. 905
Sivapragasam, C., Vincent and, P., and Vasudevan, G. (2007), Genetic programming model 906
for forecast of short and noisy Data. Hydrological Processes, 21, 266-272. 907
Solomatine, D. P., and Ostfeld, A. (2008), Data-driven modelling: Some past experiences and 908
new approaches. Journal of Hydroinformatics, 10(1), 3-22. 909
Solomatine, D. P. and Xue, Y. I. (2004), M5 model trees and neural networks: application to 910
flood forecasting in the upper reach of the Huai River in China. Journal of Hydrological 911
Engineering, 9(6), 491-501. 912
Sudheer, K. P., Gosain, A. K., and Ramasastri, K. S. (2002), A data-driven algorithm for 913
constructing artificial neural network rainfall-runoff models. Hydrological Processes, 16, 914
1325-1330. 915
Sugihara, G and May, R.M. (1990), Nonlinear forecasting as a way of distinguishing chaos 916
from measurement error in time series. Nature, 344, 734- 741. 917
Sudheer, K. P., Gosain, A. K., and Ramasastri, K. S. (2002), A data-driven algorithm for 918
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
43
constructing artificial neural network rainfall-runoff models. Hydrological Processes, 16, 919
1325-1330. 920
Theiler, J., (1986), Spurious dimension from correlation algorithms applied to limited time-921
series data. Physical Review A, 34 (3), 2427-2432. 922
Toth, E, Brath, A., and Montanari, A. (2000), Comparison of short-term rainfall prediction 923
models for real-time flood forecasting. Journal of Hydrology, 239, 132-147. 924
Venkatesan, C., Raskar, S.D., Tambe, S.S., Kulkarni, B.D., and Keshavamurty, R.N. (1997), 925
Prediction of all India summer monsoon rainfall using error-back-propagation neural 926
networks. Meteorology and Atmospheric Physics, 62 (3-4), 225-240. 927
Wang,W., Van Gelder, P.H.A.J.M., Vrijling, J.K. and Ma, J. (2006), Forecasting Daily 928
Streamflow Using Hybrid ANN Models. Journal of Hydrology, 324, 383-399. 929
Wu, C.L., Chau, K.W., and Li, Y.S., (2009), Methods to improve neural network performance 930
in daily flows prediction. Journal of Hydrology, 372(1-4), 80-93. 931
Wu, C.L., Chau, K.W. and Li, Y.S. (2008), River stage prediction based on a distributed 932
support vector regression. Journal of Hydrology, 358, 96-111. 933
Xiong, L.H., Shamseldin, A.Y., and O’Connor, K.M. (2001), A non-linear combination of 934
forecasts of rainfall-runoff models by the first-order TS fussy system for forecast of rainfall-935
runoff model. Journal of Hydrology, 245, 196-217. 936
Yates, D.N., Warner, T.T., and Leavesley, G.H. (2000), Prediction of a flash flood in complex 937
terrain. Part II: A comparison of flood discharge simulations using rainfall input from radar, a 938
dynamic model, and an automated algorithmic system. Journal of Applied Meteorology, 39 939
(6), 815-825. 940
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
44
Yu, D.L., Gomm, J.B., and Williams, D. (2000), Neural model input selection for a MIMO 941
chemical process. Engineering Applications of Artificial Intelligence, 13(1), 15-23. 942
Zhang, B., and Govindaraju, R. S. (2000), Prediction of watershed runoff using Bayesian 943
concepts and modular neural networks. Water Resources Research, 36(3), 753-762. 944
945
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
45
Figure Captions 946
Figure1. Location of Daning river basin (Map of Chongqing in the left panel, and Daning 947
watershed in the dashed box) 948
Figure 2. Rainfall series of (a) Wuxi, (b) India, (c) Zhenwan, and (d) Zhongxian 949
Figure 3. Flow chart of hard forecasting using a modular model 950
Figure 4. Implementation framework of forecast models with/without data preprocessing 951
Figure 5. Plots of ACF and PACF of the rainfall series with the 95% confidence bounds (the 952
dashed lines), (a) and (c) for Wuxi, and (b) and (d) for Zhenwan 953
Figure 6. Check of robustness of K in KNN method for (a) Wuxi, (b) India, (c) Zhenwan, 954
and (d) Zhongxian. 955
Figure 7. Singular Spectrum as a function of lag using various window lengths L for (a) Wuxi, 956
(b) India, (c) Zhenwan, and (d) Zhongxian 957
Figure 8. Sensitivity analysis of singular Spectrum on varied for (a) Wuxi, (b) India, (c) 958
Zhenwan, and (d) Zhongxian 959
Figure 9. Plots of CCF between each RC and the raw rainfall data for Zhenwan 960
Figure 10. Supervised procedure for a forecast model with SSA 961
Figure 3. Scatter plots and hyetographs of one-step-ahead forecast using ANN and MANN 962
for Wuxi ((a) and (c) from ANN, and (b) and (d) from MANN) 963
Figure 4. Scatter plots and hyetographs of one-step-ahead forecast using ANN and MANN 964
for India ((a) and (c) from ANN, and (b) and (d) from MANN) 965
Figure 5. CCFs at three forecast horizons for various lags in time of observed and forecasted 966
rainfall of ANN and MANN: Wuxi (left column) and India (right column) 967
Figure 6. Performance of ANN with SSA1 in terms of RMSE as a function of p ( L ) 968
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
46
selected RCs at various prediction horizons: (a) Wuxi and (b) India 969
Figure 7. Scatter plots and hyetographs of one-step-ahead forecast using ANN and MANN 970
with SSA2 for Wuxi ((a) and (c) from ANN, and (b) and (d) from MANN) 971
Figure 8. Scatter plots and hyetographs of one-step-ahead forecast using ANN and MANN 972
with SSA2 for India ((a) and (c) from ANN, and (b) and (d) from MANN) 973
Figure 17. CCFs at three forecast horizons using ANN and MANN with Wuxi and India ((a) 974
and (c) for Wuxi and (b) and (d) for India) 975
Figure 18. Hyetographs (detail; from 900th to 950th) of Wuxi: (1) the raw rainfall series and (2) 976
reconstructed rainfall series for one-step prediction of ANN. 977
Figure 19. CCF between inputs and output for ANN and ANN-SSA2 at three prediction 978
horizons using the Wuxi data 979
980
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
47
Table captions 981
Table 1. Pertinent information for four watersheds and the rainfall data 982
Table 2. Comparison of methods to determine mode inputs using ANN model 983
Table 3. Performance comparison of ANN with different data-transformed methods 984
Table 4. Model performances at three forecasting horizons under normal mode 985
Table 5. Model performances of ANN-MA using the Wuxi data 986
Table 6. Multiple-step predictions for Wuxi and India series using LR, K-NN, and ANN with 987
PCA1 988
Table 7. Multiple-step predictions for Wuxi and India series using LR, K-NN, and ANN 989
with PCA2 990
Table 8. Optimal p RCs for model inputs at various forecasting horizons 991
Table 9. Model performances at three forecasting horizons using MANN and ANN with the 992
SSA2 993
Table 10. RMSE of the LR model coupled with SSA2 using various L 994
995
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
48
0 15 30 45 60 75 km
Chongqing
Gaolou
Jianlou
Changan
XujiabaXining
Wuxi
Wanzhou
Yunyan
KaixianFengjie
Wushan
YangtzeRi
ver
Gaolou
Jianlou
Changan
XujiabaXining
Wuxi
Wushantown or cityhydrology stationrain gauge
996
Figure 1. 997
998
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
49
999
1000
1000 2000 3000 4000 5000 6000 70000
50
100
150
200
Time (1/1/1988-31/12/2007)
Rai
nfal
l (m
m/d
ay)
500 1000 1500 2000 2500 3000 35000
50
100
150
200
Time (1/1/1989-31/12/1998)
Rai
nfal
l (m
m/d
ay)
200 400 600 800 1000 1200 1400 16000
1000
2000
3000
4000
Time (1/1871-12/2007)
Rai
nfal
l (m
m/m
onth
)
100 200 300 400 500 6000
200
400
600
Time (1/1956-12/2007)
Rai
nfal
l (m
m/m
onth
)
Wuxi
Linear fitting
Zhenwan
Linear fitting
India
Linear fitting
Zhongxian
Linear fitting
(d)(c)
(b)(a)
1001
Figure 2. 1002
1003
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
50
1004
Figure 3. 1005
1006
Final output
( 1)Ft T mx
Raw input series
x
FCM clustering
Sub-model 1
Determination of model
inputs
Sub-model 2
Sub-model 3
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
51
1007
1008
Figure 4. 1009
Raw input series
x
LCA
AMI
PMI
FNN
CI
SLR
MOGA
Methods for model inputs
MA
PCA
SSA
MA
PCA
SSA
LR
K-NN
ANN
MANN
MA
PCA
SSA
MA
PCA
SSA
Forecasting models
Method for data preprocessing
( 1)Ft T mx
Model output
( 1)Ft T mx
( 1)Ft T mx
( 1)Ft T mx
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
52
1010
1011
0 5 10 15 20-0.2
0.2
0.6
1
AC
F
Wuxi
0 5 10 15 20-0.2
0.2
0.6
1
AC
F
Zhenwan
0 5 10 15 20-0.2
0.2
0.6
1
Lag (day)
PA
CF
0 5 10 15 20-0.2
0.2
0.6
1
Lag (day)
PA
CF
(c) (d)
(a) (b)
1012 Figure 5. 1013
1014
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
53
1015
10 20 30 4010
11
12
13
K
Val
i-RM
SE
Wuxi
10 20 30 40200
220
240
260
280
300
K
Val
i-RM
SE
India
10 20 30 4010
11
12
13
14
15
K
Val
i-RM
SE
Zhenwan
10 20 30 4052
54
56
58
K
Val
i-RM
SE
Zhongxian(d)(c)
(b)(a)
K=m+1minimum
K=m+1minimum
minimum
minimum
K=m+1
K=m+1
1016
Figure 6. 1017
1018
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
54
1 3 5 7 96
6.5
7
7.5ln
( s
ingu
lar
valu
e )
Mode
Wuxi
1 3 5 7 96.2
6.4
6.6
6.8
7
7.2
ln (
sin
gula
r va
lue
)
Mode
Zhenwan
1 3 5 7 99
10
11
12
ln (
sin
gula
r va
lue
)
Mode
India
1 3 5 7 97
7.5
8
8.5
9
ln (
sin
gula
r va
lue
)
Mode
Zhongxian
(a) (b)
(c) (d)
L=10
L=3 L=3
L=10
L=3
L=10
L=3
L=10
1019
Figure 7. 1020
1021
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
55
1022
1023
1024
1 2 3 4 56.3
6.8
7.3
ln (
sin
gula
r va
lue
)
singular number, L
Wuxi
=1=3=5=7=10
1 2 3 4 5 6 7
6.3
6.7
7.1
ln (
sin
gula
r va
lue
)
singular number, L
Zhenwan
=1=3
=5=7=10
1 2 3 4 5 6 79
10
11
12
ln (
sin
gula
r va
lue
)
singular number, L
India
=1=3=5=7=10
1 2 3 4 5 67
8
9
ln (
sin
gula
r va
lue
)
singular number, L
Zhongxian
=1=3=5=7=10
(d)(c)
(b)(a)
1025
Figure 8. 1026
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
56
1027
0 2 4 6 8 10 120
0.5
1RC1
CC
F
0 2 4 6 8 10 12-1
0
1 RC2
CC
F
0 2 4 6 8 10 12-1
0
1 RC3
CC
F
0 2 4 6 8 10 12-1
0
1 RC4
CC
F
0 2 4 6 8 10 12-1
0
1RC5
CC
F
0 2 4 6 8 10 12-0.5
0
0.5RC6
CC
F
0 2 4 6 8 10 12-0.5
0
0.5RC7
Lag (day)
CC
F
0 2 4 6 8 10 120
0.5
1Average over all RCs
Lag (day)
CC
F
1028 Figure 9. 1029
1030
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
57
1031
1032
Figure 10. 1033
1034
Choose (τ, L) for singular spectrum analysis on raw rainfall data
SSA Decomposition
Decompose raw rainfall data into L reconstructed components (RCs)
Obtain correlation coefficient matrix by correlation analysis between RC and raw rainfall data at varied lags (or termed prediction horizons) from 1 to T
Compute average correlation coefficient by averaging over L correlation coefficients at each same lag, as shown in the last plot of Figure 9.
For a specified lag (or prediction horizon), all L values of CCFs at this lag are sorted in a ascending (or ascending) order if the average CCF is negative (or positive) when the pruning algorithm for the filter of RCs is adopted
ANN is repeatedly trained and tested by new time series which is obtained by pruning operation on sorted RCs where the number of p varies from L at the beginning to 1 at the end. The optimal p is associated with the best performance of ANN.
Correlation Coefficients Sort
Reconstructed Components Filter
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
58
1035
0 50 100 1500
50
100
150
Observed
Fore
cast
ed
ANN
900 910 920 930 940 9500
50
100
150
Time (day)
Rai
nfal
l (m
m)
0 50 100 1500
50
100
150
Observed
Fore
cast
ed
MANN
900 910 920 930 940 9500
50
100
150
Time (day)
Rai
nfal
l (m
m)
observedforecasted
observedforecasted
RMSE: 8.2CE: 0.50PI: 0.60
RMSE: 10.6CE: 0.17PI: 0.32
(b)
(d)(c)
(a)
1036 Figure 11. 1037
1038
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
59
1039
1040
1041
0 500 1000 1500 2000 2500 30000
1000
2000
3000
Observed
Fore
cast
ed
ANN
150 160 170 180 190 2000
1000
2000
3000
4000
Time (month)
Rai
nfal
l (m
m)
0 500 1000 1500 2000 2500 30000
1000
2000
3000
Observed
Fore
cast
ed
MANN
150 160 170 180 190 2000
1000
2000
3000
4000
Time (month)
Rai
nfal
l (m
m)
observedforecasted
observedforecasted
(a)
(c) (d)
(b)
RMSE: 245.2CE: 0.93PI: 0.86
RMSE: 243.3CE: 0.93PI: 0.86
1042 Figure 12. 1043
1044
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
60
1045
-10 -5 0 5 100
0.5
1
Lag (day)
CC
F
ANN (Wuxi)
-10 -5 0 5 100
0.5
1
Lag (day)
CC
F
MANN (Wuxi)
-10 -5 0 5 10-1
-0.5
0
0.5
1
Lag (month)
CC
F
ANN (India)
-10 -5 0 5 10-1
-0.5
0
0.5
1
Lag (month)
CC
F
MANN (India)
One-day
Two-day
Three-day
One-day
Two-day
Three-day
One-month
Two-month
Three-month
One-month
Two-month
Three-month
(b)
(d)(c)
(a)
1046 Figure 13. 1047
1048
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
61
1049
1050
1051
5 4 3 2 14
5
6
7
8
9
10
11
12
p
RM
SE
(m
m)
Wuxi
one steptwo stepthree step
7 6 5 4 3 2 1160
180
200
220
240
260
280
300
p
RM
SE
(m
m)
India
one steptwo stepthree step
(a) (b)
1052 Figure 14. 1053
1054
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
62
0 50 100 1500
50
100
150
Observed
Fore
cast
edANN-SSA2
900 910 920 930 940 9500
50
100
150
Time (day)
Rai
nfal
l (m
m)
0 50 100 1500
50
100
150
Observed
Fore
cast
ed
MANN-SSA2
900 910 920 930 940 9500
50
100
150
Time (day)
Rai
nfal
l (m
m)
observedforecasted
observedforecasted
(a)
(c) (d)
(b)
RMSE: 4.43CE: 0.84PI: 0.87
RMSE: 3.63CE: 0.90PI: 0.92
1055 Figure 15. 1056
1057
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
63
1058
1059
1060
0 500 1000 1500 2000 2500 30000
1000
2000
3000
Observed
Fore
cast
ed
ANN-SSA2
150 160 170 180 190 2000
1000
2000
3000
4000
Time (month)
Rai
nfal
l (m
m)
0 500 1000 1500 2000 2500 30000
1000
2000
3000
Observed
Fore
cast
ed
ANN-SSA2
150 160 170 180 190 2000
1000
2000
3000
4000
Time (month)
Rai
nfal
l (m
m)
observedforecasted
observedforecasted
RMSE: 144.2CE: 0.98PI: 0.95
RMSE: 164.7CE: 0.98PI: 0.95
(b)
(d)(c)
(a)
1061 Figure 16. 1062
1063
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
64
1064
1065
1066
-10 -5 0 5 10-0.5
0
0.5
1
Lag (day)
CC
F
ANN-SSA2(Wuxi)
-10 -5 0 5 10-0.5
0
0.5
1
Lag (day)
CC
F
MANN-SSA2(Wuxi)
-10 -5 0 5 10-1
-0.5
0
0.5
1
Lag (month)
CC
F
ANN-SSA2(India)
-10 -5 0 5 10-1
-0.5
0
0.5
1
Lag (month)
CC
F
MANN-SSA2(India)
One-dayTwo-dayThree-day
One-dayTwo-dayThree-day
One-monthTwo-monthThree-month
One-monthTwo-monthThree-month
(a)
(c) (d)
(b)
1067 Figure 17. 1068
1069
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
65
1070
1071
900 905 910 915 920 925 930 935 940 945 9500
20
40
60
Time (day)
Rai
nfal
l (m
m)
Original Wuxi rainfall series
900 905 910 915 920 925 930 935 940 945 950
0
20
40
60
Time (day)
Rai
nfal
l (m
m)
Reconstructed Wuxi rainfall series
(a)
(b)
1072
Figure 18. 1073
1074
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
66
1075
0 5 10 150
0.2
0.4
0.6
0.8
1
Lag (day)
CC
F
One-step lead
0 5 10 150
0.2
0.4
0.6
0.8
1
Lag (day)
Two-step lead
0 5 10 150
0.2
0.4
0.6
0.8
1
Lag (day)
Three-step lead
Input series generated by SSA2 Raw input series
starting point
starting point
starting point
1076
Figure 19. 1077
1078
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
67
Table 11. Pertinent information for four watersheds and the rainfall data 1079
Watershed and datasets
Statistical parameters Watershed area and data period
µ Sx Cv Cs Xmin Xmax (mm) (mm) (mm) (mm)
Wuxi Original data 3.67 10.15 0.36 5.68 0.00 154 Area: Training 3.81 10.94 0.35 6.27 0.00 147 2 000 km2 Cross-validation 3.42 8.87 0.39 4.96 0.00 102 Data period: Testing 4.03 11.60 0.35 5.46 0.00 154 Jan., 1988- Dec., 2007 Zhenwan Original data 4.3 11.0 0.39 4.94 0.0 159 Area: Training 4.3 11.2 0.38 5.60 0.0 159 7554 km2 Cross-validation 4.7 11.2 0.42 4.22 0.0 125 Data period: Testing 4.0 10.9 0.37 4.97 0.0 133 Jan., 1989- Dec., 1998 India Original data 906.7 951.6 1.0 0.9 3.0 3460 Area:
Training 904.8 955.7 0.9 0.9 3.0 3393 all India
Cross-validation 918.2 969.5 0.9 1.0 8.0 3460 Data period:
Testing 898.9 927.4 1.0 0.9 16.0 3232 Jan., 1871- Dec., 2007 Zhongxian Original data 96.2 79.2 1.2 1.2 0.0 599 Area: Training 97.2 77.5 1.3 0.9 0.0 429 Cross-validation 98.6 86.8 1.1 1.9 0.0 599 Data period: Testing 91.8 74.9 1.2 0.8 0.0 306 Jan., 1956- Dec., 2007
1080
1081
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
68
Table 12. Comparison of methods to determine mode inputs using ANN model 1082
Watershed Methods m Effective inputsa Identified ANN
RMSE
Wuxi
LCA 1 20 The last 5 (5-5-1) 10.74
AMI 1 12 Except for Xt-10,t-9 (10-3-1) 10.91
PMIb 1 12 Xt,t-1,t-3,t-5,t-7,t-10,t-4 (7-8-1) 10.85
FNN 1 20 The last 14 (14-3-1) 11.02
CI 4 20 Nil
SLR 1 12 Xt-11,t-7,t-4,t-2,t-1,t (6-3-1) 10.94
MOGA 1 12 Xt, t-1 (2-6-1) 10.55
Zhenwan
LCA 1 20 The last 7 (7-4-1) 11.03
AMI 1 12 Except for Xt-11,t-10,t-9,t-8,t-2 (7-5-1) 10.95
PMI 1 12 Xt-4,t,t-1,t-3,t-11,t-5,t-10,t-6,t-9 (10-4-1) 10.98
FNN 1 20 Last 14 (14-3-1) 11.08
CI 3 20 Nil
SLR 1 12 Xt-11,t-7,t-6,t-3,t-1,t (6-3-1) 11.01
MOGA 1 12 Xt,t-4,t-7,t-9,t-11 (5-8-1) 10.43
India
LCA 1 20 the last 12 (12-5-1) 256.22
AMI 1 12 the last 12 (12-5-1) 256.22
PMI 1 12 Xt-11,t-10,t-5,t (4-5-1) 275.06
FNN 1 20 the last 5 (5-9-1) 286.04
CI 4 20 nil
SLR 1 12 except for Xt-4 (11-9-1) 258.13
MOGA 1 12 Xt-11,t-9,t-7,t-5,t-4,t-1,t (7-1-1) 277.57
Zhongxian
LCA 1 20 the last 13 (13-3-1) 51.70
AMI 1 12 Xt-11,t-10,t-6,t-5,t-4,t (6-5-1) 54.67
PMI 1 12 Xt-11,t,t-9,t-7,t-7 (5-9-1) 55.39
FNN 1 20 the last 4 (4-7-1) 59.78
CI 3 20 nil
SLR 1 12 Xt-11,t-7,t-6,t-5,t-3,t (6-6-1) 55.47
MOGA 1 12 Xt-11,t-10,t-6,t-3,t (5-2-1) 53.93
Note:a for the convenience of writing down effective inputs, “Xt, t-1”stands for Xt, Xt-1; b effective inputs 1083 from PMI are in descending order of priority. 1084
1085
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
69
Table 13. Performance comparison of ANN with different data-transformed methods 1086
Watershed Data
Transformation RMSE CE
1 2 3 a 1 2 3 Wuxi Std_raw 10.77 11.54 11.62b 0.14 0.01 0.00 Norm_raw 10.57 11.49 11.59 0.17 0.02 0.00 Std_nth_root 11.00 12.02 12.10 0.10 -0.07 -0.09 Norm_nth_root 11.15 12.01 12.09 0.08 -0.07 -0.09 Zhenwan Std_raw 11.03 11.11 11.16 0.03 0.02 0.01 Norm_raw 10.72 11.06 11.14 0.09 0.03 0.02 Std_nth_root 11.25 11.68 11.75 -0.01 -0.09 -0.10 Norm_nth_root 11.34 11.70 11.74 -0.02 -0.09 -0.09 Wuxi Std_raw 256.22 250.51 249.46 0.92 0.93 0.93 Norm_raw 251.74 246.48 250.99 0.93 0.93 0.93
Std_nth_root 259.81 253.42 256.43 0.92 0.93 0.92
Norm_nth_root 252.75 251.95 259.00 0.93 0.93 0.92
Zhongxian Std_raw 54.26 54.23 53.91 0.48 0.48 0.48 Norm_raw 52.91 53.10 52.78 0.50 0.50 0.51 Std_nth_root 52.15 53.44 53.17 0.52 0.49 0.50 Norm_nth_root 52.27 53.37 54.30 0.51 0.49 0.48
a Numbers of “1, 2, and 3” denote one-, two-, and three-day-ahead forecasting; 1087 b Result is average over 10 best runs from total 20 runs; 1088
1089
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
70
Table 14. Model performances at three forecasting horizons under the normal mode 1090
Watershed Model RMSE CE PI
1* 2* 3* 1 2 3 1 2 3
WuXi
Naïve 12.2 16.0 16.5 0.05 -0.61 -0.72 0.00 0.00 0.00
LR 10.9 11.9 12.0 0.12 -0.05 -0.07 0.28 0.41 0.43
K-NN 11.8 12.4 12.6 -0.03 -0.14 -0.18 0.16 0.36 0.38
ANN 10.6 11.5 11.6 0.17 0.02 0.00 0.32 0.45 0.47
MANN 8.2 9.2 9.0 0.50 0.38 0.40 0.60 0.65 0.68
Zhenwan
Naïve 12.4 12.2 13.6 -0.53 -0.49 -0.84 0.00 0.00 0.00
LR 11.1 11.3 11.4 0.02 -0.02 -0.03 0.39 0.42 0.45
K-NN 12.7 12.7 12.8 -0.27 -0.28 -0.30 0.21 0.28 0.31
ANN 10.7 11.1 11.1 0.09 0.03 0.02 0.43 0.45 0.48
MANN 7.9 9.6 9.9 0.50 0.27 0.23 0.69 0.59 0.59
India
Naïve 643.1 1084.5 1399.2 0.52 -0.37 -1.28 0.00 0.00 0.00
LR 286.8 301.6 302.7 0.90 0.89 0.89 0.80 0.92 0.95
K-NN 246.6 257.3 251.2 0.93 0.92 0.93 0.85 0.94 0.97
ANN 245.2 245.9 247.2 0.93 0.93 0.93 0.86 0.95 0.97
MANN 243.3 241.8 244.4 0.93 0.93 0.93 0.86 0.95 0.97
Zhongxian
Naïve 75.7 91.9 109.2 -0.03 -0.51 -1.13 0.00 0.00 -0.02
LR 56.1 57.7 58.4 0.44 0.41 0.39 0.46 0.60 0.71
K-NN 55.0 56.0 57.2 0.46 0.44 0.42 0.48 0.63 0.72
ANN 52.5 54.4 54.3 0.51 0.48 0.48 0.52 0.65 0.75
MANN 50.3 50.2 53.2 0.55 0.55 0.50 0.56 0.70 0.76 * The number of “1, 2, and 3” denote one-, two-, and three-step-ahead forecasts 1091
1092
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
71
1093
Table 15. Model performances of ANN-MA using the Wuxi data 1094
Prediction horizons
Performance index
Window length (k) for MA 1 2 3 4 5 6 7 8 9 10
One-step RMSE 10.60 10.70 10.63 10.72 10.77 10.70 10.83 10.83 10.72 10.68 CE 0.17 0.15 0.16 0.15 0.14 0.15 0.13 0.13 0.15 0.15 PI 0.32 0.31 0.32 0.31 0.30 0.31 0.29 0.29 0.31 0.31 Two-step RMSE 11.50 11.51 11.51 11.50 11.53 11.56 11.55 11.47 11.45 11.45 CE 0.02 0.02 0.02 0.02 0.01 0.01 0.01 0.02 0.03 0.03 PI 0.45 0.44 0.44 0.45 0.44 0.44 0.44 0.45 0.45 0.45 Three-step RMSE 11.60 11.58 11.58 11.55 11.61 11.58 11.56 11.51 11.52 11.52 CE 0.00 0.00 0.00 0.01 0.00 0.00 0.01 0.02 0.01 0.01 PI 0.47 0.47 0.47 0.48 0.47 0.47 0.48 0.48 0.48 0.48
1095
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
72
Table 6. Multiple-step predictions for Wuxi and India series using LR, K-NN, and ANN 1096
with PCA1 1097
Watershed Performance
index V (%)1
LR K-NN ANN 1* 2* 3* 1 2 3 1 2 3
Wuxi RMSE 85 11.0 11.9 12.0 12.1 12.5 12.7 10.6 11.5 11.6 90 11.0 11.9 12.0 12.1 12.5 12.7 10.7 11.5 11.6 95 10.9 11.9 12.0 11.8 12.4 12.6 10.6 11.5 11.6 100 10.9 11.9 12.0 11.8 12.4 12.6 10.6 11.5 11.6 CE 85 0.12 -0.05 -0.07 -0.04 -0.15 -0.19 0.16 0.01 0.00 90 0.12 -0.05 -0.07 -0.04 -0.15 -0.19 0.16 0.02 0.00 95 0.12 -0.05 -0.07 -0.03 -0.14 -0.18 0.16 0.02 0.00 100 0.12 -0.05 -0.07 -0.03 -0.14 -0.18 0.17 0.02 0.00 PI 85 0.27 0.41 0.43 0.12 0.34 0.36 0.32 0.44 0.47 90 0.27 0.41 0.43 0.12 0.34 0.36 0.31 0.45 0.47 95 0.28 0.41 0.43 0.16 0.36 0.38 0.32 0.45 0.47 100 0.28 0.41 0.43 0.16 0.36 0.38 0.32 0.45 0.47 India RMSE 85 410.9 320.6 457.7 275.9 262.7 281.1 250.4 256.1 247.0 90 294.9 307.8 311.1 260.3 256.2 276.8 248.1 252.3 249.0 95 291.6 305.0 304.6 255.3 256.5 265.0 247.7 254.2 251.4 100 286.8 301.6 302.7 246.6 257.3 251.2 245.2 245.9 247.2 CE 85 0.81 0.89 0.78 0.91 0.91 0.90 0.93 0.92 0.93 90 0.90 0.89 0.89 0.92 0.91 0.91 0.93 0.93 0.93 95 0.90 0.89 0.89 0.92 0.91 0.91 0.93 0.92 0.93 100 0.90 0.89 0.89 0.93 0.92 0.93 0.92 0.92 0.93 PI 85 0.61 0.92 0.90 0.81 0.93 0.96 0.85 0.94 0.97 90 0.79 0.92 0.95 0.83 0.94 0.96 0.85 0.95 0.97 95 0.80 0.92 0.95 0.84 0.94 0.96 0.85 0.95 0.97 100 0.80 0.92 0.95 0.85 0.94 0.97 0.84 0.94 0.97
Note: * “1, 2, and 3” denote one-, two-, and three-step-ahead forecasts; 1 “V” stands for the percentage of 1098 total variance. 1099
1100
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
73
Table 16. Multiple-step predictions for Wuxi and India series using LR, K-NN, and 1101
ANN with PCA2 1102
Watershed Performance
index V (%)1
LR K-NN ANN 1* 2* 3* 1 2 3 1 2 3
Wuxi RMSE 85 10.8 11.6 11.6 11.5 12.2 12.7 10.6 11.5 11.6 90 10.8 11.6 11.6 11.5 12.2 12.7 10.6 11.5 11.6 95 10.9 11.9 12.0 11.8 12.4 12.6 10.6 11.5 11.6 100 10.9 11.9 12.0 11.8 12.4 12.6 10.6 11.5 11.6 CE 85 0.14 0.01 0.00 0.02 -0.11 -0.19 0.16 0.02 0.00 90 0.14 0.01 0.00 0.02 -0.11 -0.19 0.16 0.02 0.00 95 0.12 -0.05 -0.07 -0.03 -0.14 -0.18 0.17 0.02 0.00 100 0.12 -0.05 -0.07 -0.03 -0.14 -0.18 0.16 0.02 0.00 PI 85 0.30 0.44 0.47 0.20 0.37 0.37 0.32 0.44 0.47 90 0.30 0.44 0.47 0.20 0.37 0.37 0.32 0.45 0.47 95 0.28 0.41 0.43 0.16 0.36 0.38 0.32 0.45 0.47 100 0.28 0.41 0.43 0.16 0.36 0.38 0.32 0.45 0.47 India RMSE 85 482.9 413.0 800.6 254.7 248.9 253.1 241.7 247.2 249.8 90 352.8 357.8 600.6 253.1 249.4 250.4 245.0 247.9 247.8 95 326.5 331.9 376.1 247.8 246.1 246.1 243.5 249.0 244.8 100 286.8 301.6 302.7 246.6 257.3 251.2 247.6 252.3 247.1 CE 85 0.73 0.80 -2.74 0.92 0.93 0.93 0.93 0.93 0.93 90 0.86 0.85 0.58 0.93 0.93 0.93 0.93 0.93 0.93 95 0.88 0.87 0.84 0.93 0.93 0.93 0.93 0.93 0.93 100 0.90 0.89 0.89 0.93 0.92 0.93 0.93 0.93 0.93 PI 85 0.44 0.86 -1.75 0.84 0.95 0.97 0.86 0.95 0.97 90 0.70 0.89 0.81 0.85 0.95 0.97 0.86 0.95 0.97 95 0.74 0.91 0.93 0.85 0.95 0.97 0.86 0.95 0.97 100 0.80 0.92 0.95 0.85 0.94 0.97 0.85 0.95 0.97
1103
1104
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
74
Table 17. Optimal p RCs for model inputs at various forecasting horizons 1105
Watershed Model Prediction horizons
Supervised filter (SSA1) Unsupervised filter (SSA2)
Optimal p RCs RMSE Optimal p RCs RMSE
Wuxi LR 1 1* 2* 6.01 1 2 6.01 2 1 5 7.73 1 5 7.73 3 1 8.40 1 8.40 K-NN 1 1 8.02 2 3 7.17 2 1 8.41 2 4 8.03 3 1 9.99 2 9.69 ANN 1 1 2 3 4.43 1 2 3 4.43 2 1 5 5.82 1 5 5.57 3 1 6.42 1 6.25 Zhenwan LR 1 1 2 3 7.19 1 2 3 7.19 2 1 7 2 7.99 1 2 7 7.99 3 1 5 8.81 1 5 8.81 K-NN 1 1 2 3 4 5 9.84 3 6 7.64 2 1 9.72 3 6 8.95 3 1 10.33 2 5 10.24 ANN 1 1 2 3 4 5.55 5 6 7 5.02 2 1 7 2 5.84 3 7 5.51 3 1 5 6.58 3 7 5.56 India LR 1 2 1 3 4 185.95 1 2 3 4 185.95 2 1 2 237.85 1 2 237.85 3 3 4 2 7 1 299.14 1 2 6 287.16 K-NN 1 2 1 3 4 5 236.90 1 2 5 6 236.39 2 1 2 247.44 1 2 5 242.65 3 3 4 2 7 249.43 1 2 5 6 243.86 ANN 1 2 166.58 1 7 164.70 2 1 2 7 173.57 1 2 7 166.30 3 3 4 2 7 1 235.59 1 5 7 172.56 Zhongxian LR 1 1 2 40.06 1 2 40.06 2 1 2 6 41.87 1 6 39.44 3 3 6 2 4 1 58.29 1 5 41.53 K-NN 1 1 2 51.78 1 2 51.78 2 1 2 53.86 1 2 3 53.32 3 3 6 2 54.15 2 3 53.16 ANN 1 1 2 3 38.39 1 5 6 34.09 2 1 44.49 1 5 6 39.45 3 3 6 46.14 1 5 34.94
Note:* the numbers of “1, 2” stand for RC1 and RC2, and these numbers in the SSA1 column is in a 1106 descending order of CCFs shown in Figure 6.12. 1107
1108
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
75
Table 18. Model performances at three forecasting horizons using MANN and ANN with the 1109
SSA2 1110
Watershed Model RMSE CE PI
1 2 3 1 2 3 1 2 3
WuXi ANN 4.43 5.57 6.25 0.84 0.78 0.70 0.87 0.88 0.84 MANN 3.63 4.32 3.93 0.90 0.86 0.89 0.92 0.92 0.94 Zhenwan ANN 5.02 5.51 5.56 0.81 0.75 0.71 0.88 0.85 0.84 MANN 3.18 3.20 3.31 0.92 0.92 0.91 0.95 0.95 0.95 India ANN 164.7 166.3 172.6 0.97 0.97 0.97 0.95 0.98 0.99 MANN 144.2 145.1 157.4 0.98 0.98 0.97 0.95 0.98 0.99 Zhongxian ANN 34.09 39.45 34.94 0.84 0.71 0.83 0.84 0.80 0.92 MANN 28.58 32.24 32.69 0.86 0.82 0.82 0.86 0.88 0.91
1111
1112
Journal of Hydrology, Vol. 389, No. 1-2, 2010, pp. 146-167
76
1113
Table 19. RMSE of the LR model coupled with SSA2 using various L 1114
Watershed Prediction horizons
L in SSA 3 4 5 6 7 8 9 10
Wuxi 1 6.13 5.94 6.01 6.41 5.83 5.81 5.61 5.51 2 7.79 7.62 7.73 8.14 7.71 7.75 7.62 7.66 3 11.76 8.61 8.40 9.04 9.23 8.72 8.56 8.62 Zhenwan 1 7.74 7.49 7.31 7.29 7.19 6.99 7.08 6.84 2 10.27 8.42 9.04 8.19 7.99 7.67 7.43 7.61 3 11.28 10.66 10.06 9.33 8.81 8.91 9.42 9.28
1115
1116